Digital Addiction

Hunt Allcott, Matthew Gentzkow, and Lena Song*

March 7, 2022

Abstract

Many have argued that digital technologies such as smartphones and social media are addictive. We
develop an economic model of digital addiction and estimate it using a randomized experiment. Tem-
porary incentives to reduce social media use have persistent effects, suggesting social media are habit
forming. Allowing people to set limits on their future screen time substantially reduces use, suggesting
self-control problems. Additional evidence suggests people are inattentive to habit formation and par-
tially unaware of self-control problems. Looking at these facts through the lens of our model suggests
that self-control problems cause 31 percent of social media use.

JEL Codes: D12, D61, D90, D91, I31, L86, O33.

Keywords: Habit formation, projection bias, self-control, temptation, naivete, commitment devices,
randomized experiments, social media.

* Allcott: Microsoft Research and NBER. [email protected]. Gentzkow: Stanford University and NBER.
[email protected]. Song: New York University. [email protected]. We thank Dan Acland, Matthew Levy, Peter Maxted,
Matthew Rabin, Dmitry Taubinsky, and seminar participants at the Behavioral Economics Annual Meeting, the Berkeley-Chicago
Behavioral Economics Workshop, Bocconi, Boston University, Chicago Harris, Columbia Business School, Cornell, Di Tella Uni-
versity, the Federal Trade Commission Microeconomics Conference, Harvard, HBS, London Business School, London School
of Economics, the Marketplace Innovation Workshop, Microsoft Research, MIT, the National Association for Business Eco-
nomics Tech Economics Conference, the New York City Media Seminar, the New York Fed, NYU, Paris School of Economics,
Princeton, Stanford Institute for Theoretical Economics, Trinity College Dublin, University of British Columbia, University Col-
lege London, USC, Wharton, and Yale for helpful comments. We thank Michael Butler, Zong Huang, Zane Kashner, Uyseok
Lee, Ana Carolina Paixao de Queiroz, Houda Nait El Barj, Bora Ozaltun, Ahmad Rahman, Andres Rodriguez, Eric Tang, and
Sherry Yan for exceptional research assistance. We thank Chris Karr and Audacious Software for dedicated work on the Phone
Dashboard app. We are grateful to the Sloan Foundation for generous support. The study was approved by Institutional Re-
view Boards at Stanford (eProtocol #50759) and NYU (IRB-FY2020-3618). This experiment was registered in the American
Economic Association Registry for randomized control trials under trial number AEARCTR-0005796; the pre-analysis plan
is available from https://www.socialscienceregistry.org/trials/5796. Replication files and survey instruments are available from
https://sites.google.com/site/allcott/research. Disclosures: Gentzkow does paid consulting work for Amazon, has done litigation
consulting for clients including Facebook, and has been a member of the Toulouse Network for Information Technology, a research
group funded by Microsoft. Both Allcott and Gentzkow are unpaid members of Facebook’s 2020 Election Research Project.

1
1 Introduction
Digital technologies occupy a large and growing share of leisure time for people around the world. The
average person with internet access spends 2.5 hours each day on social media, and there are now 3.8 billion
social media users (Kemp 2020). In a 57-country survey, people now say they spend more time consuming
online media than they do watching television (Zenith Media 2019). Americans check their smartphones 50
to 80 times each day (Deloitte 2018; Vox 2020; New York Post 2017).
A natural interpretation of these facts is that digital technologies provide tremendous consumer surplus.
However, an increasingly popular alternative view is that habit formation and self-control problems—what
we call “digital addiction”—play a substantial role. Many argue that smartphones, video games, and social
media apps may be harmful and addictive in the same ways as cigarettes, drugs, or gambling (Alter 2018;
Newport 2019; Eyal 2020). The World Health Organization (2018) has listed digital gaming disorder as an
official medical condition. Recent experimental studies find that social media use can decrease subjective
well-being (e.g. Mosquera et al. 2019; Allcott, Braghieri, Eichmeyer, and Gentzkow 2020). Figure 1 shows
that social media and smartphone use are two of the top five activities that a sample of Americans think
they do “too little” or “too much.” Compared to the other three top activities ordered at left (exercise,
retirement savings, and healthy eating), digital self-control problems have received much less attention from
economists.1
The nature and magnitude of digital addiction matter for a number of important questions. Should people
take steps to limit the amount of time they and their children spend on their smartphones and social media?
What is the best way to design digital self-control tools? How can companies that make video games, social
media, and smartphones best align their products with consumer welfare? Are proposed regulations such as
the Social Media Addiction Reduction Technology (SMART) Act a good idea?2
In this paper, we formalize an economic model of digital addiction, use a randomized experiment to
provide model-free evidence and estimate model parameters, and use the model to simulate the effects of
habit formation and self-control problems on smartphone use. We focus on six apps that account for much of
smartphone screen time and that participants report to be especially tempting: Facebook, Instagram, Twitter,
Snapchat, web browsers, and YouTube. We refer to these apps as “FITSBY.”
Our model follows Gruber and Köszegi (2001), Gul and Pesendorfer (2007), Bernheim and Rangel
(2004), and others in defining addiction as the combination of two key forces: habit formation and self-
control problems. As in Becker and Murphy (1988), habit formation means that today’s consumption
increases tomorrow’s demand. As in Laibson (1997) and others, self-control problems mean that people
consume more today than they would have chosen for themselves in advance. These two forces are central
to classic addictive goods such as cigarettes, drugs, and alcohol.
1 Among many important examples, see Charness and Gneezy (2009) and Carrera et al. (2021) on exercise, Madrian and Shea
(2001) and Carroll et al. (2009) on retirement savings, and Sadoff, Samek, and Sprenger (2020) on healthy eating.
2 This bill, introduced in 2019 by Republican Senator Josh Hawley, proposed to prohibit the use of design features such as

infinite scroll and autoplay believed to make social media more addictive, and to require companies to default users into a limit of
30 minutes per day of social media use. See Hawley (2019).

2
 Our model allows for projection bias (Loewenstein, O’Donoghue, and Rabin 2003), where people
choose as if they are inattentive to habit formation, as well as naivete about self-control problems. As in
Becker and Murphy (1988), people who perceive at least some habit formation would reduce consumption if
they know the price will increase in the future, while projection bias would dampen that effect. As in many
other models (see Ericson and Laibson 2019), people who are at least partially aware of self-control prob-
lems might want commitment devices to restrict future consumption, and people who are at least partially
unaware will underestimate future consumption.
For our experiment, we used Facebook and Instagram ads to recruit about 2,000 American adults with
Android smartphones and asked them to install Phone Dashboard, an app designed for our experiment that
records smartphone screen time and allows participants to set screen time limits. Participants completed
four surveys at three-week intervals—a baseline (survey 1) and three follow-ups (surveys 2, 3, and 4)—
that included survey measures of smartphone addiction and subjective well-being as well as predictions of
future FITSBY use. Participants answered three text message survey questions per week and kept Phone
Dashboard installed for six weeks after survey 4.
We independently randomized two treatments. The bonus treatment was a temporary subsidy of $2.50
per hour for reducing FITSBY use during the three weeks between surveys 3 and 4. We informed people
whether or not they were assigned to the bonus treatment in advance, on survey 2. The limit treatment
made available screen time limit functionality in Phone Dashboard. Participants in this group could set
personalized daily time limits for each app on their phone, with changes effective the next day. These limits
forced participants to stop using the relevant app and in most cases could not be immediately overridden,
unlike the flexible limits in existing tools such as Android’s Digital Wellbeing and iOS’s Screen Time. The
surveys encouraged participants to set limits in line with their self-reported ideal screen time, but doing so
was entirely optional. We used multiple price lists (MPLs) to elicit participants’ valuations of the bonus
treatment and the limit functionality.
The bonus treatment had persistent effects that are consistent with habit formation. The bonus reduced
FITSBY use by 56 minutes per day during the three weeks when the incentives were in effect, a 39 percent
reduction from the control group average. In the three weeks after the incentive had ended, the bonus
treatment group still used 19 minutes less per day. In the three weeks after that, they used 12 minutes less
per day.
Participants correctly predict habit formation: the effects of the bonus on predicted post-incentive
FITSBY use line up closely with the effects on actual use. However, in the three weeks between when
the bonus was announced and when it took effect, there was only a modest (and possibly zero) anticipatory
response, which is only 12 percent of what our model would predict for forward-looking habit formation
without projection bias. These results are consistent with a form of projection bias in which consumers are
aware of habit formation while consuming as if they are inattentive to it.3
3 This distinction between awareness and attention raises interesting questions about other evidence of projection bias. For

example, Busse et al. (2015) find that people are more likely to buy a convertible on sunny days. On sunny days, do people have

3
 We also find clear evidence that people have self-control problems and are at least partly aware of
them. The limit treatment reduced FITSBY screen time by 22 minutes per day (16 percent) over 12 weeks.
The effects decline slightly over the course of the experiment; this decline is consistent with some loss of
motivation, but the fact that the decline is slight means that the effects are unlikely to be driven by confusion
or temporary novelty. Although the experiment offered no incentive to set limits, 78 percent of participants
set binding limits and continued using them through the final weeks of the experiment. This far exceeds
takeup of almost all commitment devices studied in the literature reviewed by Schilbach (2019, Table 1).
On average, participants were willing to give up $4.20 for three weeks of access to the limit functionality,
and when trading off the bonus versus a fixed payment, 24 percent said they valued the bonus more highly
because they wanted to give themselves an incentive to reduce consumption. These distinct measures of
commitment demand are correlated with each other and with survey measures of addiction and desire to
reduce screen time.
Notwithstanding their demand for commitment, participants seem to slightly underestimate their self-
control problems. The control group modestly but repeatedly underestimated their future FITSBY use in
all of our surveys, even though use is fairly steady over time and we reminded them of recent past use
before asking them to predict. On average, the control group underestimated next-period FITSBY use by
6.1 minutes per day, or about 4 percent.
To further evaluate whether our interventions reduced addiction in a way that participants perceive to
be beneficial, we examine effects on a variety of survey outcomes. On both the main surveys and text mes-
sages, the bonus and limit treatments significantly reduced an index of smartphone addiction adapted from
the psychology literature. For example, both treatment groups reported being less likely to use their phone
longer than intended, use their phone to distract from anxiety or fall asleep, have difficulty putting down
their phone, lose sleep from phone use, procrastinate by using their phone, and use their phone mindlessly.
Both treatment groups reported improved alignment between ideal and actual screen time. The bonus treat-
ment group also scored higher on an index of subjective well-being, with statistically significant increases
in components related to concentration and avoiding distraction and statistically insignificant changes in
measures of happiness, life satisfaction, anxiety, and depression. Finally, both treatments are well-targeted
in the sense that effects were more positive for people who report more interest in reducing their use and
who score higher on our addiction measures at baseline.
In the final section of the paper, we look at these results through the lens of our structural model. The
model allows us to translate our short-run experimental estimates into effects on long-run steady state be-
havior, to quantify the magnitude of the effects we observe in terms of economically meaningful parameters,
and to decompose the role of different behavioral forces through counterfactuals. We first estimate the model
parameters by matching key moments from the experiment. We model the limit treatment as eliminating
share ω of self-control problems, and for our primary estimates we conservatively assume ω = 1. The es-
timates reflect our experimental results: substantial habit formation and self-control problems, substantial
different beliefs about future weather or how much they would drive a convertible?

4
projection bias, and slight naivete about self-control problems. We then evaluate how steady-state consump-
tion would change in counterfactuals where we eliminate self-control problems. Without habit formation, a
conservative estimate of the effect of self-control problems is the effect of giving people screen time limit
functionality: 22 minutes per day. But habit formation amplifies the effect of self-control problems, as the
increase in current consumption also increases future marginal utility. In the presence of habit formation,
our primary model prediction is that eliminating self-control problems would reduce FITSBY use by 48
minutes per day, or 31 percent of baseline use. Alternative assumptions mostly imply more self-control
problems, more attention to habit formation, and larger effects on use.
Our results should be interpreted with caution for several reasons. First, our experiment took place dur-
ing the beginning of the coronavirus pandemic. Our survey evidence suggests that this increased screen time
but did not have clear effects on the magnitude of self-control problems. Furthermore, even as the pandemic
evolved over the three-month experiment, average screen time and the treatment effects of the limit were
fairly stable. Second, our estimates apply to the 2,000 people who selected into our experiment, and these
people are not representative of U.S. adults. When we reweight our estimates to more closely approximate
national average demographic characteristics, the modeled effect of self-control problems increases. Third,
our model’s predictions of FITSBY use without self-control problems depend on assumptions such as linear
demand and geometric decay of habit stock. Fourth, our analysis is partial equilibrium in the sense that
we do not model network effects and other externalities across users. If one person’s social media use in-
creases others’ use, such positive network externalities would magnify the effects of self-control problems
on population-wide social media use. Finally, our surveys walked participants through a process of set-
ting optional screen time limits that implemented their self-reported ideal screen time, and we hypothesize
that simply offering time limit functionality without walking through that process would have had smaller
effects.4
Our work builds on several existing literatures. We extend a distinguished literature documenting present
focus in diverse settings including exercise, healthy eating, consumption-savings decisions, and laboratory
tasks (Ericson and Laibson 2019).5 Ours is one of a small handful of papers that estimate the parameters of a
present focus model with partial naivete using field (instead of laboratory) behavior.6 The digital self-control
4 While Carrera et al. (2021) show that takeup of commitment devices can be driven by experimenter demand effects or decision-

making noise instead of perceived self-control problems, there are three reasons why their concerns are less likely to apply to our
experiment. First, while Carrera et al. (2021) studied one-time takeup of an unfamiliar commitment contract, our participants re-
peatedly set and continually kept screen time limits over a 12-week period. Second, we estimate even larger perceived self-control
problems using participants’ valuations of the bonus treatment, which leverages an alternative methodology favored by Carrera
et al. (2021) as well as Acland and Levy (2012), Augenblick and Rabin (2019), Chaloupka, Levy, and White (2019), Allcott, Kim,
Taubinsky, and Zinman (2021), and Strack and Taubinsky (2021). Third, unlike Carrera et al. (2021), we find strong correlations
between use of screen time limits and other measures of perceived self-control problems.
5 This includes Read and Van Leeuwen (1998), Fang and Silverman (2004), Shapiro (2005), Shui and Ausubel (2005), Ashraf,

Karlan, and Yin (2006), DellaVigna and Malmendier (2006), Paserman (2008), Gine, Karlan, and Zinman (2010), Duflo, Kremer,
and Robinson (2011), Acland and Levy (2012), Andreoni and Sprenger (2012a; 2012b), Augenblick, Niederle, and Sprenger (2015),
Beshears et al. (2015), Goda et al. (2015), Kaur, Kremer, and Mullainathan (2015), Laibson et al. (2015), Royer, Stehr, and Sydnor
(2015), Exley and Naecker (2017), Augenblick (2018), Kuchler and Pagel (2018), Toussaert (2018), Augenblick and Rabin (2019),
Casaburi and Macchiavello (2019), Schilbach (2019), John (2019), Toussaert (2018), and Sadoff, Samek, and Sprenger (2020).
6 To our knowledge, these are Allcott, Kim, Taubinsky, and Zinman (2021), Bai et al. (2018), Carrera et al. (2021), Chaloupka,

5
problems we study are particularly interesting because this is one of the few domains where market forces
have created commitment devices, such as blockers for smartphone apps, email, and websites (Laibson
2018). Our results suggest additional unmet demand for these commitment devices.
We also extend a distinguished literature on habit formation. One set of papers documents persistent
impacts of temporary interventions in settings such as academic performance, energy use, exercise, hand
washing, political protest, smoking, recycling, voting, water use, and weight loss.7 We provide evidence in
an important new domain. A second set of papers tests for forward-looking habit formation using belief elic-
itation or advance responses to future price changes, sometimes interpreting such forward-looking behavior
as support for “rational” models of addiction.8 We estimate anticipatory responses using an experimental
approach that, like the one in Hussam et al. (2019), addresses many confounds that arise in observational
data (Chaloupka and Warner 1999; Gruber and Köszegi 2001; Auld and Grootendorst 2004; Rees-Jones and
Rozema 2020). Furthermore, we use our model to actually estimate the magnitude of projection bias, which
is important because earlier studies that reject a null hypothesis of fully myopic habit formation could still
be consistent with substantial projection bias.
Finally, we extend three literatures that speak directly to digital addiction. The first literature includes
theoretical papers modeling temptation in digital networks (Makarov 2011; Liu, Sockin, and Xiong 2020).
The second includes experimental papers studying the effects of social media use on outcomes such as sub-
jective well-being and academic performance.9 The third studies the effects of digital self-control tools.10
Hoong (2021) is particularly related, and is an important antecedent to our study. In a smaller-scale exper-
iment, she pioneers the use of encouragement to adopt self-control tools, compares predicted and ideal use
to actual use, and shows results consistent with significant self-control problems. Our paper helps to unify
the previous empirical literature with a formal model of digital addiction, relatively large sample, multiple
treatment arms that convincingly identify habit formation and self-control problems using several different
strategies, and robust measurement of screen time and survey outcomes.
Section 2 sets up the model. Sections 3–5 detail the experimental design, data, and model-free results.
Section 6 presents the model estimation strategy and parameter estimates, and Section 7 presents the mod-
eled effects of temptation on time use.

Levy, and White (2019), and Skiba and Tobacman (2018).

7 This includes Gerber, Green, and Shachar (2003), Charness and Gneezy (2009), Gine, Karlan, and Zinman (2010), Ferraro,

Miranda, and Price (2011), John et al. (2011), Allcott and Rogers (2014), Bernedo, Ferraro, and Price (2014), Acland and Levy
(2015), Royer, Stehr, and Sydnor (2015), Fujiwara, Meng, and Vogl (2016), Levitt, List, and Sadoff (2016), Beshears and Milkman
(2017), Brandon et al. (2017), Carrera et al. (2018), Allcott, Braghieri, Eichmeyer, and Gentzkow (2020), Bursztyn et al. (2020),
Gosnell, List, and Metcalfe (2020), and Van Soest and Vollaard (2019).
8 This includes Chaloupka (1991), Becker, Grossman, and Murphy (1994), Gruber and Köszegi (2001), Acland and Levy (2015),

Hussam et al. (2019), and Do and Jacoby (2020).

9 This includes Sagioglu and Greitemeyer (2014), Tromholt (2016), Hunt et al. (2018), Vanman, Baker, and Tobin (2018),

Mosquera et al. (2019), Allcott, Braghieri, Eichmeyer, and Gentzkow (2020), and Collis and Eggers (2019).
10 This includes Marotta and Acquisti (2017) and Acland and Chow (2018).

6
2 Model
The goal of the model is to formalize the meaning of “digital addiction” and foreshadow how we identify
the model parameters using our experiment.
In each period t ≤ T , consumers choose consumption of a good xt sold at price pt that delivers flow
utility ut (xt ; st , pt ). To model habit formation, utility depends on a stock st of past consumption that evolves
according to
st+1 = ρ (st + xt ) , (1)

where ρ ∈ [0, 1) captures the strength of habit formation. Habit formation captures why temporary price
changes generate persistent effects in our experiment.
To model self-control problems, we follow Banerjee and Mullainathan (2010) in modeling x as a tempta-
tion good. Before period t, consumers consider period t flow utility to be ut (xt ; st , pt ). In period t, however,
consumers choose as if period t flow utility is ut (xt ; st , pt ) + γxt , where γ ≥ 0 reflects the amount of temp-
tation. If γ > 0, consumers choose more xt in period t than they would choose in advance. This temptation
good framework generates similar predictions to the quasi-hyperbolic model from Laibson (1997) and Gru-
ber and Köszegi (2001), but it naturally matches our application to a single addictive good and yields simpler
estimating equations where temptation is additively separable.
Consumers may misperceive temptation: before period t, consumers predict that in period t, they will
consider flow utility to be ut (xt ; st , pt ) + γ̃xt . We say that consumers are fully naive if γ̃ = 0, and fully
sophisticated if γ̃ = γ. Partial naivete captures why our experiment participants underestimate xt when
asked to predict in advance. Partial sophistication captures why our participants want commitment devices
to change their future behavior.
Following Loewenstein, O’Donoghue, and Rabin (2003), we allow the possibility of projection bias,
in which consumers choose as if to maximize a weighted average of utility given the current habit stock
st and utility given the predicted habit stock s̃r in future period r > t. We let α denote the weight on the
current habit stock, and thus the magnitude of projection bias. Projection bias captures why consumers in
our experiment might not reduce consumption in anticipation of a known future price change. We assume
that consumers are fully naive about projection bias; sophistication would introduce strategic incentives to
adjust current consumption to offset future bias.11
Following O’Donoghue and Rabin (1999) and others, we solve for perception-perfect equilibrium strate-
gies, where consumers maximize current utility given predictions of future behavior. Let xt (st , γ, pt ) denote
a strategy of the period-t self, which depends on habit stock, temptation, and the vector of future prices
pt = {pt , pt+1 , ..., pT }. Let x̃r (sr , γ̃, p r ) be a consumer’s prediction, as of period t < r, of her period-r
11 Loewenstein, O’Donoghue, and Rabin (2003, page 1219) also assume naivete about projection bias, writing that “because
this time inconsistency derives solely from misprediction of future utilities, it would make little sense to assume that the person
is fully aware of it.” We note that our formulation of projection bias is slightly different than in Loewenstein, O’Donoghue, and
Rabin (2003): while their consumers’ predictions of future consumption are biased due to projection bias, our consumers predict
consumption accounting for habit formation, but choose as if they are inattentive to it. This matches our empirical results.

7
strategy. A strategy profile (x0∗ , ..., xT∗ ) is perception perfect if in each period t
" #
α ∑Tr=t+1 δ r−t ur (x̃r∗ (st , γ̃, p r ) ; st , pr )
xt∗ (st , γ, pt ) = arg max ut (xt ; st , pt ) + γxt + , (2)
xt +(1 − α) ∑Tr=t+1 δ r−t ur (x̃r∗ (s̃r , γ̃, p r ) ; s̃r , pr )

where δ ≤ 1 is the discount factor.

Predicted and actual consumption differ due to naivete about temptation and projection bias and the re-
sulting misprediction of habit stock. We assume that the equilibrium prediction x̃r∗ (sr , γ̃, p r ) is the solution to
equation (2) with α = 0 and γ = γ̃. Predicted habit stock s̃r evolves according to s̃r+1 = ρ (s̃r + x̃r∗ (s̃r , γ̃, p r )).12
The “rational” habit formation model of Becker and Murphy (1988) is the special case with α = 0 and
γ̃ = γ = 0.
To estimate the model, we follow Becker and Murphy (1988) and Gruber and Köszegi (2001) in spe-
cializing to the case of quadratic flow utility:

η 2
ut (xt ; st , pt ) = x + ζ xt st + φ st + (ξt − pt )xt (3)
2 t
where η < 0 measures the demand slope, ζ regulates the extent of habit formation, φ is the direct effect of
habit stock on utility (which could be positive or negative), and ξt is a deterministic period-specific demand
shifter. This can be microfounded by assuming that consumers have income w that they must spend in each
period, and income not spent on xt is spent on a numeraire ct = w − pt xt that is additively separable in ut . In
this specification, ut is in units of dollars per period.

3 Experimental Design
3.1 Overview
Our experiment is designed to provide direct evidence on the magnitude of habit formation, perceived habit
formation, temptation, and perceived temptation, as well as to identify the remaining key parameters of the
quadratic model. The experiment ran from March 22 to July 26, 2020, with participants completing an
intake questionnaire and four surveys. Figure 2 summarizes the experimental design, and Table 1 presents
sample sizes at each step.
Between March 22 and April 8, we recruited participants using Facebook and Instagram ads. Appendix
Figure A1 presents the ads. To minimize sample selection bias, the ads did not hint at our research questions
or suggest that the study was related to smartphone use or social media. 3, 271, 165 unique users were shown
one of the ads, of whom 26,101 clicked on it. This 0.8 percent click-through rate is close to the average
click-through rate on Facebook ads (Irvine 2018).
12 Sincethe predicted equilibrium strategy x̃r∗ (sr , γ̃, p r ) conditions on the state sr inherited at time r, it will be the same when
evaluated in all periods t < r. However, the predicted action in period r is not generally the same when evaluated in all periods
t < r, as the predicted s̃r will differ by t.

8
 Clicking on the ad took the participant to a brief screening survey, which included several background
questions, the consent form, and instructions on how to install Phone Dashboard. To be eligible, participants
had to be a U.S. resident between 18 and 64 years old, use an Android as their primary phone, and use only
one smartphone regularly. 18,589 people satisfied these criteria, of whom 8,514 consented to participate in
the study. Of these, 5,320 successfully installed Phone Dashboard and finished the intake survey.
Surveys 1–4 were administered on Sundays at three week intervals between April 12th and June 14th.
We define t = 1, 2, 3, ... to be the three-week periods beginning Monday April 13th, so period t is the three
weeks immediately after survey t. For our data analysis and interventions, we want to exclude survey days,
so all periods are 20 days long, from a Monday to a Saturday. Survey 1 recorded participant demographics.
We describe the other survey content below.
As illustrated in Figure 2, we randomized participants into bonus and limit treatment conditions (detailed
below) using a factorial design. We randomized participants to the Bonus, Bonus Control, or the Multiple
Price List (MPL) group with 25, 75, and 0.2 percent probability, respectively. We independently randomized
participants to the Limit or Limit Control groups with 60 and 40 percent probability, respectively. We refer
to the intersection of the Bonus Control and Limit Control groups as the Control group. We balanced the
randomization within eight strata defined by above- versus below-median baseline FITSBY use, restriction
index, and addiction index (described below). The treatments began on survey 2.
All participants received $5 for completing the baseline survey and $25 if they completed the remaining
surveys and kept Phone Dashboard installed through July 26th. Participants were also entered in a drawing
for a $500 gift card, in which two winners were drawn.
As shown in Table 1, 4,038 participants completed survey 1. We dropped 1,912 of these participants
from the experiment after survey 1 because they reported that they already used another app to limit their
phone use (5 percent of the sample) or failed data quality checks.13 The remaining 2,126 participants were
invited to take survey 2, of whom 2,053 opened the survey and reached the point where the treatments began.
Of those, 1,938 completed the study—remarkably low attrition for a 12-week study with multiple surveys.
In addition to back-loading the survey payments, several other factors contributed to our limited attrition.
There were two surveys (the intake and survey 1) before the treatments began, inducing likely attriters to
attrit beforehand. At the beginning of survey 2, just before the treatments began, we informed people that
“anyone who drops out after this page can really damage the entire study,” and offered them a choice to drop
out at that moment or commit to finishing the whole study. For participants who had not yet completed each
of surveys 2–4, we sent daily reminders for six days after the survey had been fielded, and after four days we
began offering an additional payment for completing all remaining surveys. We also sent reminder emails
to people who had failed to respond to two consecutive text messages.
13 Participantsfailed data quality checks if they (i) did not to promise to “provide my best answers” on our surveys; (ii) reported
having idiosyncratic bugs with Phone Dashboard; (iii) failed to answer more than two of our text message questions between survey
1 and survey 2; (iv) had a device that was incompatible with Phone Dashboard; or (v) were missing screen time data during the
baseline period.

9
3.2 Phone Dashboard
Phone Dashboard is an Android app that was developed by a company called Audacious Software for our
experiment. Appendix Figure A2 presents screenshots. Our experiment includes only Android users because
a similar functionality cannot be implemented by third-party apps on iOS.
Phone Dashboard records the app that is in the foreground of a smartphone every five seconds when
the screen is on; we use these data to construct our measure of consumption. It does not record the content
that the user is viewing within the app. Users can see their cumulative screen time by day and by week
on the Phone Dashboard home screen. This usage information was designed to be particularly useful for
participants in the Bonus and Limit groups who might want to track their usage, but the Control group also
used the app: the Bonus, Limit, and Control groups used Phone Dashboard for an average of 1.4, 1.5 and
1.0 minutes per day during periods 2–5.

3.3 Bonus Treatment

The bonus treatment was designed to identify projection bias (the parameter α), actual habit formation (ρ
and ζ ), and the curvature of utility (η). To facilitate the multiple price list (MPL) described below, survey 2
explained the bonus to all participants before telling them whether they were selected to receive it and when
it would be in force. Participants were told,
If you’re selected for the Screen Time Bonus, you would receive $50 for every hour you reduce your
average daily FITSBY screen time below a Bonus Benchmark of [X] hours per day over the 3-week period,
up to $150.
The survey then gave several examples, including:

• If you reduce your FITSBY screen time to $[X-1] hours and 30 minutes per day over the next 3 weeks,
you would receive $25.

• If your FITSBY screen time is above $X hours per day, you would receive $0.

We set the Bonus Benchmark [X] as the participant’s average FITSBY hours per day from period 1, rounded
up to the nearest integer.
After the MPL described below, the Bonus group was informed that they had been randomly selected
to receive the bonus for screen time reductions during period 3—i.e., starting in three weeks. The Bonus
Control group was informed that they would not receive the bonus. To ensure that participants understood,
each participant had to answer a question by correctly indicating their bonus treatment condition before
advancing. We also sent three text messages reminders to the Bonus group during period 2, which read
“Don’t forget, we’ll pay you $50 for every hour you reduce your average daily screen time between May 24
and June 14. There is no bonus for changing your screen time before then.” People were asked to respond
to the text message to confirm that they had read it. Survey 3 included an additional reminder for the Bonus
treatment group. While we received substantial feedback on the surveys and many emails from our 2,000

10
participants during the study and our earlier pilots, none of these interactions suggested confusion about the
timing of the bonus.
The Bonus group’s anticipatory response to the bonus in period 2 (before the incentive was in effect)
provides information about the magnitude of projection bias α. The contemporaneous response in period
3 (while the incentive was in effect) provides information about the price response parameter η. The long-
term effects in periods 4 and 5 (after the incentive had ended) provides information about the magnitude and
decay of habit (ζ and ρ).

3.4 Limit Treatment

The limit treatment was designed to understand self-control problems and help identify the temptation pa-
rameter γ. The Limit treatment group was given access to functionality in Phone Dashboard that allows
users to set daily time limits for each app on their phone; see Appendix Figure A2 for screenshots. Any
changes to the limits take effect the next day. Phone Dashboard serves five-minute and one-minute push no-
tifications as an app’s daily time limit approaches. When the limit arrives, users can “snooze” their limit and
get an additional amount of time that they specify—but starting only after a delay. Within the Limit group,
we randomly assigned participants with equal probability to delays of 0, 2, 5, or 20 minutes or a condition
where the ability to snooze was disabled. To keep the scope of this paper manageable, we focus only on the
comparison between the Limit and Limit Control groups; we plan to study the variation in snooze delays in
a separate paper. To reduce attrition and uninstallation, Phone Dashboard also allows people to permanently
opt out of the limits; about 4 percent of the Limit group did so.
The Limit group was first given access to the Phone Dashboard limit functionality on survey 2, after
the Screen Time Bonus multiple price list described below, and they retained access to the feature for the
duration of the experiment. To introduce the limits, we first gave participants instructions on how to set daily
app usage limits for themselves. The survey then asked participants what time limits they would like to set
for themselves on each FITSBY app over the next three weeks. We then asked participants to update their
Phone Dashboard app, which activated the limit functionality, and encouraged them to set the limits they
had reported a moment earlier. The Limit Control group was never told about limits and continued to have
a version of Phone Dashboard that did not have the limit functionality.
In the analysis below, we interpret use of the limits as evidence of perceived self-control problems
(γ̃ > 0).

3.5 Bonus and Limit Valuations

We used incentive-compatible multiple price list mechanisms to elicit valuations of the Screen Time Bonus
and the limit functionality. Because both the bonus and the limit functionality reduce future social media
use, these valuations help identify perceived temptation γ̃.
All multiple price lists included a table with a series of choices between “Option A” and “Option B”

11
in separate rows. Option B was the same in each row, while Option A included an amount of money
that decreased monotonically from top to bottom. Participants would typically choose Option A at the top
and Option B at the bottom, and we infer their valuation of Option B from the row where they switch. To
encourage valid answers, participants who did not switch between Option A and Option B exactly once were
alerted to this fact and given a chance to change their answers. All MPLs were incentivized, as described
below. To help participants become familiar with MPLs, survey 1 included an incentivized practice MPL
that asked participants to choose between receiving different survey completion payments at different times.
Our approach to valuing the Screen Time Bonus builds on Allcott, Kim, Taubinsky, and Zinman (2021)
and Carrera et al. (2021). Survey 2 informed participants of their average daily FITSBY screen time over
the past three weeks and asked them to predict their screen time over the next three weeks. The survey
then introduced the Screen Time Bonus and asked participants to predict how much they would reduce their
FITSBY screen time relative to their original prediction if they were selected for the bonus.
After these two predictions, we asked participants to make a hypothetical choice between the Screen
Time Bonus and a payment equal to their expected earnings from the bonus. The survey described potential
considerations as follows:

• You might prefer $[expected earnings] instead of the Screen Time Bonus if you don’t want any pressure
to reduce your screen time.

• You might prefer the Screen Time Bonus instead of $[expected earnings] if you want to give yourself
extra incentive to use your phone less.

Participants then completed an MPL where Option B was receiving the Screen Time Bonus, and Option A
was receiving a payment ranging from $150 to $0.
To make the MPL incentive compatible, participants were told, “Last week, the computer randomly
selected some participants to receive what they choose on the multiple price list below, and also randomly
selected one of the rows to be ‘the question that counts.’ If you were randomly selected to participate, you
will be paid based on what you choose in that row.” 0.2 percent of participants were randomly assigned to
the MPL group that received what they chose on a randomly selected row.
On survey 3, the Limit group completed an MPL that elicited valuations of the Phone Dashboard limit
functions. Option B was retaining access to the Phone Dashboard limit functions, and Option A was having
those functions disabled for the following three weeks in exchange for a dollar payment that ranged from
$20 to -$1. The MPL group received what they chose on a randomly selected row.

3.6 Predicted Use

At the end of surveys 2, 3 and 4, we elicited predictions of future FITSBY use. These predictions help
identify the degree of naivete or sophistication about temptation—the difference between γ and γ̃.
Before each elicitation, we told each participant their average FITSBY screen time over the previous
three weeks. Surveys 2 and 3 also reminded the Bonus and Limit groups about the bonus and limits. Survey

12
2 then elicited predictions of FITSBY screen time for the next three weeks (period 2), the three weeks after
that (period 3), and the three weeks after that (period 4). Survey 3 elicited separate predictions for periods
3, 4, and 5. Survey 4 elicited separate predictions for periods 4 and 5.
Predictions were incentivized. Survey 2 told participants, “Answer carefully, because you might earn a
Prediction Reward. After the study ends, we will pick a prediction question at random and check how close
your prediction is. If your predicted daily screen time is within 15 minutes of your actual screen time, we
will pay you an additional $X.” We randomized the prediction reward X to be $1 or $5, each with 50 percent
probability.

3.7 Survey Outcome Variables

Surveys 1, 3, and 4 asked questions designed to measure participants’ perceptions of their addiction and
subjective well-being (SWB). For the nine weeks between survey 1 and survey 4, we also sent three text
messages per week with a subset of questions that we thought were important to ask in real time instead
of retrospectively. Using these questions, we construct five pre-specified outcome variables. Appendix A.1
presents details on the survey questions.
Ideal use change. The survey said,

Some people say they use their smartphone too much and ideally would use it less. Other people are
happy with their usage or would ideally use it more. How do you feel about your smartphone use over the
past 3 weeks?

• I use my smartphone too much.

• I use my smartphone the right amount.

• I use my smartphone too little.

For people who said they used their smartphone “too much” or “too little,” we then asked, Relative to your
actual use over the past 3 weeks, by how much would you ideally have [reduced/increased] your smartphone
use? The ideal use change variable is the answer to this question, in percent.
Addiction scale. Our addiction scale is a battery of 16 questions modified from two well-established
survey scales, the Mobile Phone Problem Use Scale (Bianchi and Phillips 2005) and the Bergen Facebook
Addiction Scale (Andreassen et al. 2012). The questions attempt to measure the six core components of
addition identified in the addiction literature: salience, tolerance, mood modification, relapse, withdrawal,
and conflict (Griffiths 2005).
The survey asked, In the past three weeks, how often have you ..., with a matrix of 16 questions, such as

• used your phone longer than intended?

• felt anxious when you don’t have your phone?

13
 • lost sleep due to using your phone late at night?

Possible answers were Never, Rarely, Sometimes, Often, and Always, which we coded as 0, 0.25, 0.5, 0.75,
and 1, respectively. Addiction scale is the sum of these numerical scores for the 16 questions.
SMS addiction scale. The SMS addiction scale includes shortened versions of nine questions from the
addiction scale. Examples include:

• In the past day, did you feel like you had an easy time controlling your screen time?

• In the past day, did you use your phone mindlessly?

• When you woke up today, did you immediately check social media, text messages, or email?

People were instructed to text back their answers on a scale from 1 (not at all) to 10 (definitely). SMS
addiction scale is the sum of these scores for the nine questions.
Phone makes life better. The survey asked, To what extent do you think your smartphone use makes
your life better or worse? Responses were on a scale from -5 (“Makes my life worse”) through 0 (“Neutral”)
to +5 (“Makes my life better”).
Subjective well-being. We use standard measures from the subjective well-being literature, mostly fol-
lowing the measures from our own earlier work (Allcott, Braghieri, Eichmeyer, and Gentzkow 2020). The
survey asked,

Please tell us the extent to which you agree or disagree with each of the following statements. Over the
last three weeks, with a matrix of seven questions:

• . . . I was a happy person

• . . . I was satisfied with my life

• . . . I felt anxious

• . . . I felt depressed

• . . . I could concentrate on what I was doing

• . . . I was easily distracted

• . . . I slept well

Possible answers were on a seven-point scale from “strongly disagree” through “neutral” to “strongly agree,”
which were coded as -1, -2/3, -1/3, 0, 1/3, 2/3, and 1, respectively. The variable subjective well-being is
the sum of these numerical scores for the seven questions, after reversing anxious, depressed, and easily
distracted so that more positive reflects better subjective well-being.

14
 Indices. We define the survey index to be the sum of the five survey outcome variables described above,
weighted by the baseline inverse covariance matrix as described by Anderson (2008). When presenting
results and constructing this index, we orient the variables so that more positive values imply normatively
better outcomes. Thus, we multiply addiction scale and SMS addiction scale by (-1).
We define the restriction index to be the sum of interest in limits (with the four categorical answers
coded as 0, 1, 2, and 3) and ideal use change, after normalizing each into standard deviation units. We
define the addiction index to be the sum of addiction scale and phone makes life better after normalizing
each into standard deviation units. We use these two indices for stratified randomization and as moderators
when testing for heterogeneous treatment effects.

3.8 Pre-Analysis Plan

We submitted our pre-analysis plan (PAP) on May 4th, the day that post-treatment data collection began.
The PAP specified (i) the equation for treatment effect estimation (equation 4 below); (ii) the construction of
the survey outcome variables and indices described in Section 3.7, the limit tightness variable, and the win-
sorization of predicted FITSBY use; and (iii) the analysis of heterogeneous treatment effects by splitting the
sample on above- versus below-median values of six moderators: education, age, gender, baseline FITSBY
use, restriction index, and addiction index. The PAP also included shells of Tables 1, 2, and A1–A3, as well
as Figures 1–6, A1–A4, A8, and A28–A34.
We deviate from the PAP in five ways. First, the bottom left panel of Figure 3 includes results from
each addiction scale question, whereas the PAP figure shell presented the sum across all questions. Second,
we clarify that our analysis sample includes only the balanced panel of people who completed the study.
Results are essentially identical if we use an unbalanced panel that includes data from attriters before they
attritted, but the balanced panel is helpful in ensuring that our habit formation results are not spuriously
driven by attrition. Third, three figures from the PAP are not included here, as we plan to study them in a
separate paper. Fourth, Figure 6 includes predicted FITSBY use from all surveys before period t, whereas
the PAP figure shell presented predictions from only the survey immediately before period t. Fifth, we use
equation (4) for subgroup analysis, whereas the PAP specified that we would use an instrumental variables
regression. We present the pre-specified instrumental variables estimates in Appendix D.4. The results are
similar, and we decided that equation (4) was simpler.

4 Data
The analysis sample for all results reported below is the balanced panel of 1, 933 participants who were
assigned to either Bonus or Bonus Control (not the MPL group), completed all four surveys, and kept Phone
Dashboard installed until the end of the study on July 26. This group’s attrition rate after being informed of
treatment was (1 − 1, 933/2, 048) × 100% ≈ 5.6 percent. Attrition rates and observable characteristics are
balanced across the bonus and limit treatment conditions; see Appendix Tables A1 and A2.

15
 Table 2 quantifies the representativeness of our analysis sample on observables, by comparing their
demographics to the U.S. adult population. Our sample is more educated, more heavily female, younger,
and slightly lower-income than the U.S. population. We estimate an alternative specification of our structural
model with sample weights to adjust for these observable differences.
Table 2 also shows that the average participant had 333 minutes per day of screen time during the
baseline period, of which 153 minutes (46 percent) was on FITSBY apps. Different sources report very
different estimates of average social media use and smartphone screen time for U.S. adults, so we do not
report nationwide averages in the table. Kemp (2020) reports that internet users in the U.S. and worldwide,
respectively, spend an average of 123 and 144 minutes per day on social media, mostly on mobile devices.
Wurmser (2020) and Brown (2019) report national averages of 186 and 324 minutes of total smartphone
screen time per day, respectively. The comparisons suggest that the heavy use in our sample may not be far
from the national average.
During the baseline period, the average participant used Facebook, browsers, YouTube, Instagram,
Snapchat, and Twitter for 69, 44, 23, 24, 15, and 15 minutes per day, respectively; see Appendix Figure
A3. Appendix Figure A4 presents the distribution of baseline FITSBY use. Appendix Table A3 presents
descriptive statistics for the survey outcome variables.

5 Model-Free Results
5.1 Treatment Effect Estimating Equation
To estimate treatment effects, define Yit as an outcome for participant i for period t. Yit could represent a
survey outcome variable measured on survey t ∈ {3, 4}, or period t FITSBY use. Define Li and Bi as Limit
and Bonus group indicators. Define X i1 as a vector of baseline covariates: baseline FITSBY use and, if and
only if Y is a survey outcome variable, the baseline value Yi1 and the baseline value of survey index. Define
νit as a vector of the eight randomization stratum indicators, allowing separate coefficients for each period
t. We estimate the effects of the limit and bonus treatments using the following regression:

Yit = τtB Bi + τtL Li + βt X i1 + νit + εit . (4)

When combining data across multiple periods, we cluster standard errors by participant.

5.2 Baseline Qualitative Evidence

Figure 3 presents qualitative evidence on digital addiction from the baseline survey. The top two panels
present the variables in the restriction index. The top left panel shows that 23 percent of people reported
being “moderately” or “very” interested in setting time use limits on their smartphone apps, while 34 percent
reported being “not at all” interested. The top right panel presents the distribution of responses to the ideal
use change question. 42 percent of people said that they used their smartphone the right amount over the

16
past three weeks, and only 0.5 percent said that they used it too little. Among people who said they used
their smartphone too much, the average ideal reduction was 34 percent.
Survey 1 also asked people to report their ideal use change for specific apps or categories. FITSBY,
games, video streaming, and messaging are the nine apps on which people want to reduce screen time the
most; see Appendix Figure A8. Facebook is by far the most tempting app: the average participant would
ideally reduce Facebook use by 22 percent. The average participant did not want to change their use of
email, news, and maps and wanted to slightly increase use of phone, music, and podcast apps.
The bottom two panels present the variables in the addiction index. The bottom left panel presents the
share of participants who responded “often” or “always” on each question in the addiction scale. The top
seven questions capture three components of moderate addictions (salience, tolerance, and mood modifica-
tion); 33 percent of participants often or always experience each of these, and 84 percent often or always
experience at least one. The bottom nine questions capture three components of more severe addictions
(relapse, withdrawal, or conflict); 11 percent of participants often or always experience each of these, and
41 percent often or always experience at least one. The bottom right panel shows that while most people
think that their smartphone use makes their life better, 19 percent think that it makes their life worse. Taken
together, these results suggest substantial heterogeneity: many people report experiences consistent with
addiction, while many others do not.
Our experiment took place during the coronavirus pandemic, which significantly disrupted people’s daily
routines. To understand how this might affect our results, we included several baseline survey questions,
which we report in Appendix C. 78 percent of people reported having more free time as a result of the
pandemic, and 88 percent of people reported that the pandemic had increased their phone use. However, it
is not clear that the pandemic affected the extent of self-control problems: the means and distributions of key
qualitative measures of addiction that we also asked for 2019, ideal use change and phone makes life better,
were statistically different but economically similar. Ideal use change is closer to zero in 2020 compared to
in 2019, suggesting less perceived self-control problems, but phone makes life better is also less positive,
suggesting more perceived self-control problems.

5.3 Bonus Treatment and Habit Formation

The darker coefficients in Figure 4 present the effect of the bonus on FITSBY use, estimated using equation
(4). Recall that the bonus provides an incentive to reduce FITSBY use in period 3, but we informed partic-
ipants about whether or not they were offered the bonus at the beginning of period 2. The incentive is $50
per average hour measured over the 20-day period, or $2.50 per hour of consumption.
In period 3 (while the incentive was in effect), the Bonus group reduced FITSBY use by 56 minutes per
day, or 39 percent relative to the Control group. This is a striking price response: it implies that participants
value a substantial share of smartphone FITSBY use at less than $2.50 per hour.
In periods 4 and 5 (after the incentive had ended), the Bonus group still reduced FITSBY use by 19 and
12 minutes per day, respectively. This persistent effect suggests substantial habit formation, implying ζ > 0

17
in our model. The decay of the effect in period 5 relative to period 4 provides information about the habit
stock decay parameter ρ in our model.
In period 2 (before the incentive was in effect), the Bonus group reduced FITSBY use by 5.1 min-
utes per day, which is marginally statistically significant. This is consistent with the model’s prediction
that a consumer who perceives habit formation should reduce period 2 consumption in order to reduce pe-
riod 3 marginal utility, which makes it easier to reduce period 3 consumption in response to the financial
incentive. However, additional evidence suggests some caution about interpreting the period 2 effect as
forward-looking habit formation. Appendix Figures A9 and A10 break out the period 2 effect separately by
day and week, showing that it loads mostly on the first half of the period. If anything, forward-looking habit
formation would predict the opposite pattern, with larger anticipatory effects closer to the beginning of the
incentive period. Possible explanations include intertemporal substitution, a temporary idiosyncratic effect,
and the salience of the bonus after its introduction on survey 2.14

5.4 Limit Treatment and Temptation

The Limit group made extensive use of the limit functionality. To summarize the stringency of time limits,
we define the variable limit tightness to be the amount by which a user’s limits would have hypothetically
reduced screen time if applied to their baseline use.15 Limit tightness equals zero (instead of missing)
for an app if the participant doesn’t have the app or doesn’t set a limit, so this variable speaks to what
apps contribute the most to aggregate temptation. About 89 percent of the Limit group had positive limit
tightness at some point during the experiment, suggesting that they set binding screen time limits, and 78
had positive limit tightness in period 5, meaning that they kept those limits for more than three weeks after
the final survey. Participants most wanted to restrict Facebook, web browsers, YouTube, and Instagram:
limit tightness averaged 20, 10, 8, and 6 minutes per day on those apps, respectively, across periods 2–5.
Across all apps, the Limit group’s average limit tightness was 53 minutes per day. See Appendix Figures
A11 and A12 for details.
The lighter coefficients on Figure 4 present the effect of the limit on FITSBY use. These actual effects
are smaller than the limit tightness values in the previous paragraph primarily because users snooze the
14 Although we stratified randomization on period 1 FITSBY use and also control for period 1 use when estimating equation (4),

some idiosyncratic factor could temporarily affect consumption in Bonus versus Bonus Control at the beginning of period 2. Some
evidence supports this possibility: Appendix Figure A9 shows that consumption is slightly lower in the Bonus group compared to
Bonus Control in the final 11 days of period 1. Salience could also play a role, although as described in Section 3.3, we took many
steps to eliminate confusion about the timing of the bonus incentive period, and participants likely would have emailed our team if
they were confused.
15 Specifically, define x
iadt as the screen time of person i on app a on day d in period t. Define hiat as the average screen time
limit in place in period t, and define Nd∈t=1 as the number of days in the baseline period. Limit tightness is

1
Hiat = ∑ max {0, xiad1 − hiat } . (5)
Nd∈t=1 d∈t=1
If the daily limit hiat would not have been binding in baseline day d, the max function returns 0. If hiat would have been binding
in day d, then the max function returns the excess screen time on that day. We aggregate over apps to construct user-level limit
tightness Hit = ∑a Hiat .

18
limits. Access to the limit functionality reduced use in periods 2–5 by an average of 22 minutes per day, or
16 percent relative to the Control group. The effects attenuate only slightly as the experiment continues, and
the effect is still 19 minutes per day in the last week of period 5. This is notable because while surveys 2 and
3 walked people through a limit-setting process and survey 4 included an optional review of the limits, the
end of period 5 is nine weeks after survey 3 and six weeks after survey 4. These continuing effects suggest
that while motivation might decrease over time, use of the limits is not primarily driven by confusion or
temporary novelty. Furthermore, Appendix D.1 shows that limit tightness is correlated in expected ways
with bonus and limit valuations and with survey measures of addiction and desire to reduce screen time.
This evidence consistently points toward perceived self-control problems, implying γ̃ > 0 in our model.
When we add the interaction between Bonus and Limit group indicators to equation (4), the main effects
are similar and the interaction terms are not statistically significant; see Appendix Figure A13.

5.5 Substitution
Figure 5 presents usage effects of the bonus (in period 3 only) and the limit (across periods 2–5) separately by
app. Among the FITSBY apps, Facebook sees the largest reductions, followed by web browsers, YouTube,
Instagram, Twitter, and Snapchat. The effects on other apps (the right-most coefficients) provide evidence on
the extent to which participants substituted FITSBY time to alternative apps. The bonus has no statistically
detectable effect on use of other apps in period 3, and the confidence intervals rule out any substantial
substitution relative to the 56 minutes per day reduction in FITSBY use. The limit induces substitution of
12 minutes per day, so that roughly half of the FITSBY screen time that the limit eliminates moves to other
apps where people had been less likely to set limits.
One important limitation is that we cannot directly monitor FITSBY use on devices other than the par-
ticipant’s smartphone. We screened out potential participants who reported using more than one smartphone
regularly, but our remaining participants may still have used desktops, tablets, or other devices. To provide
some evidence on this substitution, survey 4 asked participants to estimate their FITSBY use on other de-
vices in period 3 compared to the three weeks before they joined the study. The results, shown in Appendix
Figure A14, imply that the limit increased FITSBY use on other devices by a marginally significant 4.2
minutes per day. The bonus reduced the amount of time they spent on FITSBY on other devices by 8.1
minutes per day, suggesting that time on other devices was a mild complement in this case.
The differences in substitution induced by the bonus versus limit are notable. In a simple model where
other apps and devices are either complements or substitutes for smartphone FITSBY use, the substitution
effects described above might have the same sign for both the bonus and limit and might be in proportion to
their direct effects on smartphone FITSBY use. In contrast, a much smaller share of the effect on FITSBY
use is substituted to other smartphone apps for the bonus compared to the limit, and the self-reported effects
on FITSBY use on other devices have opposite signs for the bonus versus the limit. This is an interesting
result to understand in future work.

19
5.6 Predicted versus Actual Use
Figure 6 presents predicted and actual FITSBY use in the Control condition, where participants had neither
the bonus nor the limit functionality. As specified in our pre-analysis plan, we winsorize predicted use at
no more than 60 minutes per day more or less than actual use in the corresponding period. Within each
period, the left-most spike is actual average use. The spikes to the right are average predictions. The point
estimates show that people consistently underestimate their use in all future periods, even though actual
use is fairly stable throughout the experiment and the surveys had reminded them of their past use before
eliciting predictions. This is consistent with naivete, implying γ̃ < γ in our model.
Figure 7 presents predicted versus actual habit formation. Within each period, the left-most point is the
treatment effect of the bonus on actual use, reproduced from Figure 4. Recall that before the multiple price
list for the Screen Time Bonus on survey 2, we asked people to report the percent by which they thought
the bonus would reduce their FITSBY use. Their estimates (translated into minutes using their status quo
predictions) are almost exactly correct on average: 52 minutes per day. Then on survey 3, we asked people
to predict their use in future periods. Figure 7 also presents treatment effects of the bonus on predicted use,
estimated from equation (4). The figure shows that people correctly predict that the bonus will reduce their
consumption in period 3 and that this reduction will persist even after the incentive is no longer in effect. If
anything, comparing the time path of actual versus predicted effects suggests that people overestimate the
extent of habit formation. Overall, these results suggest that people are well aware of habit formation.
Appendix D.1 presents additional results that validate that the usage predictions are meaningful. Pre-
dicted use lines up well with actual use, and the higher ($5 instead of $1) prediction accuracy reward slightly
reduces the absolute value of the prediction error but has tightly estimated zero effects on predicted use, ac-
tual use, and the level of the prediction error.

5.7 Bonus and Limit Valuations

On the survey 3 multiple price list, the average Limit group participant was willing to give up a $4.20
fixed payment for three weeks of access to the limit functionality. About 58 percent of participants were
willing to give up at least some money for the limits, and 20 percent were willing to give up more than $10;
see Appendix Figure A17. This willingness to pay for a commitment device is consistent with perceived
self-control problems (γ̃ > 0) and unmet market demand for digital self-control tools.
On the survey 2 multiple price list, people who perceive self-control problems should prefer the Screen
Time Bonus over higher fixed payments, as the incentive helps bring future use in line with current pref-
erences. We show in Appendix E.5 that participants’ average valuation of the bonus is consistent with
perceived self-control problems (γ̃ > 0).
Appendix D.1 presents additional results that validate that the MPL responses are meaningful. First,
participants’ valuations of the bonus are correlated with the amount of money they could expect to earn.
Second, the bonus and limit valuations are correlated with each other and with limit tightness, ideal use

20
change, addiction scale, SMS addiction scale, and other variables in expected ways. Third, after the bonus
MPL, we asked people to “select the statement that best describes your thinking when trading off the Screen
Time Bonus against the fixed payment.” 24 percent responded that “I wanted to give myself an incentive
to use my phone less over the next three weeks, even though it might result in a smaller payment,” and this
group had a higher average valuation.

5.8 Effects on Survey Outcomes

Figure 8 presents the effects of the bonus and limit treatments on the survey outcomes described in Section
3.7. The outcome variables are signed so more positive effects always correspond to less addiction and/or
higher subjective well-being. Following our pre-analysis plan, when estimating effects on survey outcomes,
we constrain the limit effect to be the same for surveys 3 and 4 (because we correctly anticipated similar
“first stage” effects on FITSBY use in both periods 2 and 3) and we report the bonus effect only for survey
4 (because we correctly anticipated negligible “first stage” effects on FITSBY use in period 2).16
Figure 8 shows that both interventions significantly reduced self-reported measures of addiction. Ap-
pendix Table A6 presents coefficient estimates and p-values. The bonus effect is larger than the limit effect
for five of the six variables, consistent with the bonus’s larger effects on FITSBY use. The bonus decreased
ideal use change by 0.41 standard deviations (about 9 percentage points), while the limit decreased it by 0.23
standard deviations (about 5 percentage points). Both interventions reduced addiction scale and SMS ad-
diction scale by 0.08 to 0.16 standard deviations, or about 0.21–0.44 points on the 16-point addiction scale.
Both interventions statistically significantly reduced the chance that people reported using their smartphones
to relax to go to sleep, losing sleep from use, using longer than intended, using to distract from anxiety, hav-
ing difficulty putting down their phone, using mindlessly, and other specific measures from the addiction
scales; see Appendix Figures A23 and A24. The limit treatment statistically significantly increased the
extent to which people thought their smartphone use made their life better, while the bonus did not.
The bonus and limit treatments increased subjective well-being (SWB) by 0.09 standard deviations (p ≈
0.026) and 0.04 standard deviations (p ≈ 0.18) respectively. The sharpened False Discovery Rate-adjusted
p-values (see Benjamini and Hochberg 1995) are 0.09 and 0.24, respectively. These SWB effects appear to
be driven particularly by improved concentration and reduced distraction; see Appendix Figure A25. The
effects of the bonus and limit on happiness, life satisfaction, depression, and anxiety are individually and
collectively insignificant, while the effects of the bonus (but not the limit) on concentration, distraction, and
sleep quality are collectively significant. Both interventions improved survey index, the inverse covariance-
weighted average of the five survey outcome variables, by about 0.2 standard deviations.
One point of comparison for the SWB effects is Allcott, Braghieri, Eichmeyer, and Gentzkow (2020).
They find that deactivating subjects’ Facebook accounts for a four week period increased an index of SWB
16 Appendix Figure A22 presents the treatment effects on survey outcomes separately for surveys 3 and 4. The limit effects on
surveys 3 and 4 are statistically indistinguishable. Although the bonus did not substantially affect consumption in period 2, the
Bonus group reported more ideal use reduction and more addiction on survey 3. One potential explanation is that the Bonus group
hoped to reduce FITSBY use in anticipation of the period 3 incentive, and these survey responses reflect their failure to do so.

21
by a statistically significant 0.09 standard deviations. Although the two interventions had similar effects on
time use—deactivation in Allcott, Braghieri, Eichmeyer, and Gentzkow (2020) reduced Facebook use by 60
minutes per day for 27 days, while our Screen Time Bonus reduced FITSBY use by 56 minutes per day for
20 days—they differed on a number of dimensions including the apps affected and the time period in which
the study took place.
Appendix Figure A26 presents effects on survey index in subgroups with above- and below-median val-
ues of our six pre-specified moderators. There is little heterogeneity with respect to the first four moderators,
other than that the limit seems to have larger effects on women. However, the effects of both interventions
are 2–3 times larger for people with above-median baseline values of restriction index, which measures
interest in restricting smartphone time use, and addiction index. This implies that the interventions are well-
targeted: they have larger effects for people who report wanting and needing them the most. Consistent with
this, point estimates suggest that the bonus and limit both have larger effects on FITSBY use for people with
higher restriction index and addiction index, although the differences are not as significant; see Appendix
Figure A27. This targeting result need not have been the case: for example, it could have been that more
addicted people were less likely to feel that the limit functionality worked well for them.

6 Estimating the Model

6.1 Setup
We now turn to our model to simulate the effect of temptation on steady-state FITSBY use. In the model
from Section 2, temptation and habit formation interact, because the current consumption increase caused
by temptation also increases future consumption. The long-run effect of temptation could therefore be
different than any effects identified during the experiment. In this section, we estimate the model’s structural
parameters. In the next section, we simulate steady-state FITSBY use with counterfactual self-control and
habit formation parameters.
We estimate the model using indirect inference: we derive equations that characterize how a consumer
from our model would behave in our experiment, and we solve for the structural parameters consistent with
the data. In our baseline estimates, we assume that all parameters other than ξ are homogeneous across
consumers, although we relax this assumption in an extension that allows heterogeneity in temptation γ and
perceived temptation γ̃.
In describing the estimation strategy, we focus on a “restricted model” where we set the anticipatory
bonus effect τ2B to zero. This implies full projection bias (α = 1), and thus that consumption decisions
maximize current-period flow utility with no dynamic considerations. This substantially simplifies the ex-
position and, as we will show, has little impact on the results. Appendix E presents our “unrestricted model,”
in which we use the empirical τ2B and allow partial projection bias.
In the restricted model, consumers maximize current-period flow utility from equation (3), giving equi-

22
librium consumption
ζ st + ξt − pt + γ
xt∗ (st , γ, pt ) = . (6)
−η
∂ xt∗
We define λ := ∂ st as the effect of habit stock on consumption; λ = −ζ /η in the restricted model. In a
ρ
steady state with constant s, ξ , and p, we must have sss = ρ (sss + xss ), and thus sss = 1−ρ xss .
We model the Screen Time Bonus as a price pB = $2.50 per hour in period 3 plus a fixed payment.17
We model the limit functionality as an intervention that elimates share ω of perceived and actual temptation.
We conservatively assume ω = 1 in our primary estimates, and we consider alternative assumptions below.
We assume that when predicting period t consumption on the survey at the beginning of period t, consumers
use perceived temptation γ̃ but are aware of projection bias, so the prediction is denoted xt∗ (st , γ̃, pt ).
Figure 9 illustrates temptation, naivete, and our identification strategies. The three demand curves are
desired demand xt∗ (st , 0, pt ) according to preferences before period t, predicted demand xt∗ (st , γ̃, pt ) as of
survey t, and actual demand xt∗ (st , γ, pt ). The actual equilibrium at p = 0 is point L, and the predicted
equilibrium is at point C, so the distance CL is Control group misprediction mC := xt∗ (st , γ, pt ) − xt∗ (st , γ̃, pt ).
The bonus moves the equilibrium to point J in period 3, so the contemporaneous bonus effect τ3B is the
distance JK. The limit treatment moves the equilibrium to point G, so the limit treatment effect τ L is the
distance GL.

6.2 Estimating Equations

Unlike many applications of indirect inference, we derive equations that allow us to directly solve for the
model parameters, so we do not need to use an optimization routine to search for parameters that fit the data.
We estimate the parameters in stages, as parameters estimated in the first few equations are used as inputs
to subsequent equations. We estimate confidence intervals by bootstrapping. Appendix G presents formal
derivations and additional details.

Habit Formation

We first estimate ρ from the decay of the bonus treatment effects. Taking the expectations over ξ in the
Bonus and Bonus Control groups, we can write the average treatment effect of the bonus on period 4 con-
sumption as the result of the decayed period 3 effect. Similarly, the average treatment effect in period 5
results from the cumulative decayed effects from periods 3 and 4:
17 Modeling the bonus as a linear price simplifies the model substantially, although it is an approximation: 13 percent of the

Bonus group hit the $150 payment limit because they reduced period 3 FITSBY use by more than three hours per day relative
to their Bonus Benchmark, and 3.5 percent used more than their Bonus Benchmark. These two subgroups in practice faced zero
subsidy for marginal screen time reductions, although they may not have known that.

23
 τ4B = λ ρτ3B (7)
τ5B = λ ρτ4B + ρ 2 τ3 . B


(8)

Dividing those two equations gives

τ5B τ4B
ρ= − . (9)
τ4B τ3B

This equation shows that if the bonus effect is more persistent between periods 4 and 5, we infer that habit
stock is more persistent (a larger ρ).
In the unrestricted model in Appendix E, we also estimate λ , because it is useful in estimating the other
parameters. To provide a comparison, we also estimate λ in the restricted model by rearranging equation
τ4B
(7) and inserting the ρ from equation (9): λ = ρτ3B
.

Price Response and Habit Stock Effect on Marginal Utility

After estimating ρ, we estimate η and ζ from the magnitude and decay of the bonus treatment effects. For
each of periods 3 and 4, we take the expectations over ξ of equilibrium consumption in the Bonus and Bonus
Control groups, difference the two, and rearrange, giving

pB
η= (10)
τ3B
−ητ4B
ζ= . (11)
ρτ3B

Figure 9 illustrates the first equation: the inverse demand slope η is just the ratio of pB (the vertical distance
KL) to τ3B (the horizontal distance JK). The second equation shows that if the bonus effect is more persistent
between periods 3 and 4, we infer that habit stock has a larger effect on marginal utility (a higher ζ ).

Naivete about Temptation

Next, we estimate naivete about temptation γ − γ̃ using misprediction in the Control group. To solve for
γ − γ̃, we take the expectations over ξ of actual consumption and consumption as predicted at the beginning
of the period, difference the two, and rearrange, giving

γ − γ̃ = −ηmC . (12)

24
Figure 9 illustrates: naivete γ − γ̃ is the vertical distance HC between actual and predicted marginal utility,
and this can be inferred by multiplying Control group average misprediction mC (the horizontal distance CL
between actual and predicted demand) by the inverse demand slope η.

Temptation

To estimate temptation γ, we take the expectations over ξ of equilibrium consumption in the Limit and Limit
Control groups, difference the two, and rearrange, giving

γ = ητ2L . (13)

Figure 9 illustrates: temptation γ is the vertical distance LM between desired and actual demand, and this
can be inferred by multiplying the effect of removing temptation (τ2L , the horizontal distance GL between
long-run and present demand) by the inverse demand slope η. We then substitute the estimated γ into
equation (12) to infer γ̃.

Intercept

Finally, we back out the distribution of ξ that fits the distribution of baseline consumption. We assume that
ρ
participant i’s baseline consumption xi1 was in a steady state. Substituting sss = 1−ρ xss into equilibrium
consumption from equation (6) and rearranging gives
 
ρ
ξi = p − γ + xi1 −η − ζ . (14)
1−ρ
This equation shows that we infer larger ξi for people with higher baseline consumption xi1 .
In the unrestricted model in Appendix E, equilibrium consumption also depends on φ , the direct effect of
habit stock on utility. Our data do not allow us to separately identify φ from ξ , so we estimate an “intercept”
κi := (1 − α)δ ρ(φ − ξi ) + ξi that includes both of these structural parameters. In the restricted model with
α = 1, this simplifies to κi = ξi .

6.3 Empirical Moments

Table 3 presents the moments used to estimate the restricted model. The bonus and limit effects τtB and τ2L are
as displayed in Figure 4. Control group misprediction mC is the average across periods 2–4 of the difference
between actual period t FITSBY use and the prediction for period t elicited on survey t, as displayed in
Figure 6. The unrestricted model and our robustness checks also use the anticipatory bonus effect τ2B and
additional parameters presented in Appendix Table A7. In light of the discussion in Section 5.3, we omit
the first half of period 2 when we estimate τ2B .18
18 Appendix Table A8 presents parameter estimates when we use all of period 2 to estimate τ B . The estimated projection bias α
2
is smaller, as expected, but the other parameter estimates are very similar.

25
6.4 Parameter Estimates
Table 4 presents our point estimates and bootstrapped 95 percent confidence intervals. Column 1 presents
the restricted model described above (fixing τ2B = 0 and α = 1), while column 2 presents the unrestricted
model described in Appendix E. Since the estimated τ2B is close to zero and α̂ is close to one, the estimates
in the two columns are very similar.
In column 1, we estimate λ̂ ≈ 1.15 and ρ̂ ≈ 0.299. In our model, this implies that an exogenous
consumption increase of 1 minute per day over a three week period will cause consumption to increase by
λ̂ ρ̂ ≈ 0.34 minutes per day in the next three-week period, and λ̂ ρ̂ 2 ≈ 0.10 minutes per day in the period
after that.
Consistent with the small and statistically insignificant anticipatory bonus effect τ2B in the second half
of period 2, we estimate α̂ ≈ 0.897 in the unrestricted model in column 2, which is marginally significantly
different from one. The point estimate suggests that participants were attentive to only (1 − α̂) × 100% ≈
10.3 percent of habit formation. Inserting the estimates of λ , ρ, η, and ζ into equation (24) in Appendix
E, we calculate that τ2B would have needed to be −16.1 minutes per day (compared to the actual point
estimate of −1.96 minutes per day in the second half of period 2) to estimate zero projection bias (α = 0).
In other words, the anticipatory bonus effect is only 12 percent of what our model would predict with
fully forward-looking (“rational”) habit formation. This is striking when combined with the evidence from
Figure 7 that participants correctly predicted habit formation. It is consistent with a model in which people
are intellectually aware of habit formation but consume as if they are inattentive to it.
Since the restricted model estimating equations are so simple, one can easily calculate the point estimates
in column 1 with the moments from Table 3. For example, the Control group underestimated FITSBY use
by an average of 6.13 minutes per day on surveys 2–4. Inserting that into equation (12) gives a naivete of
− γ̃ = −η̂ · mC ≈ −(−2.68) · (6.13/60) ≈ 0.274 $/hour in column 1.
γ[
The limit changed period 2 FITSBY use by −24.3 minutes per day. Inserting that into equation (13)
gives temptation γ̂ = η̂τ2L ≈ (−2.68) · (-24.3/60) ≈ 1.09 $/hour in column 1. This estimate implies that
a tax on FITSBY use of $1.09 per hour would reduce consumption to the level our participants would
− γ̃ by this γ̂ suggests that our participants
choose for themselves in advance. Dividing estimated naivete γ[
underestimate temptation by 0.274/1.09 × 100% ≈ 25 percent.
Appendix E.5 presents alternative estimates of temptation γ in the restricted and unrestricted models.
First, we infer perceived temptation using participants’ valuations of the limit functionality and the Screen
Time Bonus, following Acland and Levy (2012), Augenblick and Rabin (2019), Chaloupka, Levy, and White
(2019), Allcott, Kim, Taubinsky, and Zinman (2021), and Carrera et al. (2021). Second, we generalize the
model to include multiple temptation goods, using the self-reports of substitution to FITSBY use on other
devices discussed in Section 5.5. Third, we assume that the limit treatment eliminates share ω ∈ [0, 1] of
temptation, relaxing the assumption of ω = 1 in our primary estimates; we estimate ω from differences in
self-reported ideal use change between the Limit and Limit Control groups. Finally, we allow for individual-
specific heterogeneity in γ, using the distribution of limit tightness set by Limit group participants. These

26
alternative approaches all imply temptation γ between about $1 and $3 per hour, and our primary estimate
of $1.09 per hour is relatively conservative.

7 Counterfactuals: Effects of Temptation on Time Use

7.1 Methodology
Using the parameter estimates from the previous section, we can predict the effects of changes in temptation
and habit formation on steady-state FITSBY use. Equation (21) in Appendix E characterizes steady-state
consumption in the unrestricted model. Using that equation, we can predict participant i’s steady-state
FITSBY use at p = 0 as a function of any values of habit formation, temptation, and steady-state mispredic-
tion parameters {ζ , γ, γ̃, mss }:

ˆ
h   i
κ̂i + (1 − α̂)δ ρ̂ (ζ − η̂) mss − 1 + λ̃ γ̃ + γ
x̂i,ss (ζ , γ, γ̃, mss ) = 2 . (15)
−η̂ − (1 − α̂)δ ρ̂(ζ − η̂) − ζ ρ̂−(1−α̂)δ ρ̂
1−ρ̂

The sample average prediction is denoted x̂¯ss (ζ , γ, γ̃, mss ). As discussed in Appendix E.3, we assume that
the predicted λ̃ equals the estimated λ̂ , that steady-state misprediction mss equals observed Control group
misprediction mC , and that the discount factor is δ = 0.997 per three-week period, consistent with a five
percent annual discount rate.
Since we can’t identify φ (the direct effect of habit stock on utility), we must hold constant each par-
ticipant’s intercept κi := (1 − α)δ ρ(φ − ξi ) + ξi across counterfactuals in the restricted model. Since this
intercept contains ρ and α, we can’t predict consumption with counterfactual values of ρ or α.
In the restricted model with α = 1, equation (15) simplifies to

ξˆi + γ
x̂i,ss (ρ, γ) = , (16)
−η̂ − ζˆ 1−ρ
ρ

ρ
which could also be derived from substituting sss = 1−ρ xss into equation (6). Steady-state misprediction mss
and perceived temptation γ̃ do not affect steady-state consumption in the restricted model because consumers
simply maximize current-period flow utility.

7.2 Counterfactual Results

Figure 10 presents point estimates and bootstrapped 95 percent confidence intervals for predicted average
FITSBY use at counterfactual parameter values. For each counterfactual, we present predictions from the
restricted model (α = 1) and unrestricted model (α = α̂). We label the restricted model predictions as our
primary results, because they are simpler and more conservative.  
The first “counterfactual” is the baseline at our point estimates: x̂ss ζˆ , γ̂, γ̃ˆ, m̂C . This mechanically

27
matches
 baseline
 average FITSBY use of 153 minutes per day. The second counterfactual removes naivete:
¯x̂ss ζˆ , γ̂, γ̂, 0 .19 As described above, naivete has no effect when α = 1. Because naivete is so small and
projection bias is so strong, the point estimate with α= α̂ is very
 close to the baseline.
¯ ˆ
The third counterfactual removes temptation: x̂ss ζ , 0, 0, 0 . Relative to baseline, removing temptation
reduces predicted FITSBY use by 48 minutes per day (31 percent) with α = 1. Thus, our primary estimate
is that smartphone FITSBY use would be 31 percent lower without self-control problems.
The fourth and fifth counterfactuals remove habit formation, first with temptation and then without:
x̂¯ss 0, γ̂, γ̃ˆ, m̂C and then x̂¯ss (0, 0, 0, 0). We emphasize that habit formation on its own is not a departure

from rationality (Becker and Murphy 1988), and it could capture forces such as learning and investment that
increase consumer welfare. Relative to baseline, removing habit formation reduces predicted FITSBY use
by 75 minutes per day with α = 1. Without habit formation, the effect of removing temptation (going from
the fourth to the fifth counterfactual) is just the limit treatment effect (τ2L ≈ -24.3 minutes per day), which is
about half of the effect of removing temptation with habit formation (47.5 minutes per day with α = 1).20
This quantifies how habit formation magnifies the effects of temptation, because current temptation increases
current consumption and thus future demand.
We highlight one important tension in our results: Figure 4 shows that the limit effects decay slightly
over periods 2–5, while our model predicts that the limit effects should grow over time as the Limit group’s
habit stock diminishes. One potential explanation is that habit formation works differently in response to
prices versus self-control tools. Another potential explanation is that motivation to use the limit functionality
decays enough that it outweighs the habit stock effect.
Appendix Table A15 presents 19 alternative estimates of the effects of temptation on steady-state FITSBY
use across the restricted and unrestricted models. Consistent with the fact that our primary estimates of γ
are smaller than most alternative estimates, our primary estimates of the steady-state temptation effects are
also relatively conservative. Furthermore, weighting our sample on observables to look more like the U.S.
adult population also increases the predicted effects of temptation on consumption. This means that while
our sample may still be non-representative on unobservable characteristics, sample selection bias captured
by observables causes us to understate the effects of temptation on FITSBY use.21
Since we don’t identify φ (the direct effect of habit stock on utility), we can’t do a full welfare analysis.
The relatively elastic demand—from Section 5.3, 39 percent of consumption is worth less than $2.50 per
hour—suggests that participants do not have strong preferences over how to spend this marginal time, so the
welfare losses from self-control problems might be limited. On the other hand, even small individual-level
losses might be substantial when aggregated over many social media users. In a static model, the deadweight
19 SinceFigure 7 shows that participants predicted habit formation fairly accurately, we attribute all of steady-state misprediction
mss to naivete about temptation.
20 Without habit formation, the effect of removing temptation on x̂¯ is γ̂ , which equals τ L after substituting γ̂ = τ L η̂ .
ss −η̂ 2 2
21 Appendix Tables A11–A13 present the demographics, moments, and parameter estimates in the weighted sample. Appendix

Table A14 presents the numbers plotted in Figure 10. Appendix Figure A35 presents the distribution of modeled temptation effects
across participants, using the Limit group’s distribution of limit tightness to identify heterogeneity in temptation. The effect is less
than 10 minutes per day for 26 percent of participants, and over 100 minutes per day for 13 percent.

28
loss from temptation would be the triangle GLM on Figure 9: −τ L γ/2 ≈ −(-24.3/60) × 1.09/2 ≈ $0.22 per
day, or $4.62 per three-week period. This is closely consistent with the average valuation of $4.20 for three
weeks of access to the limit functionality. Aggregated across 240 million American social media users (Pew
Research Center 2021), this would be $4.62×(52/3) × 0.24 ≈ $19.2 billion per year in welfare losses from
overuse of social media caused by self-control problems. For comparison, Facebook’s total global profits
in 2020 were $29 billion (United States Securities and Exchange Commission 2020). However, we don’t
know how these effects would cumulate over time, as represented by φ : for example, after a longer period
of reduced screen time, people might find more peace of mind or regret the loss of online interactions with
friends and family.

8 Conclusion
While digital technologies provide important benefits, some argue that they can be addictive and harmful.
We formalize this argument in an economic model and transparently estimate the parameters using data from
a field experiment. The Screen Time Bonus intervention had persistent effects after the incentives ended,
suggesting that smartphone social media use is habit forming. Participants predicted these persistent effects
on surveys but did not reduce FITSBY use before the bonus was in effect, suggesting that they are aware of
but inattentive to habit formation. Participants used the screen time limit functionality when we offered it
in the experiment, and this functionality reduced FITSBY use by over 20 minutes per day, suggesting that
social media use involves self-control problems. The Control group repeatedly underestimated future use,
suggesting slight naivete. Many participants reported indicators of smartphone addiction on surveys, and
both the bonus and limit interventions reduced this self-reported addiction. Looking at these facts through
the lens of our economic model implies that self-control problems magnified by habit formation might be
responsible for 31 percent of social media use. These results suggest that better aligning digital technologies
with well-being should be an important goal of users, parents, technology workers, investors, and regulators.
Our results raise many additional questions; here are two. First, what are the underlying mechanisms
and microfoundations that generate the persistent bonus treatment effects? We model this persistence simply
through a capital stock of past consumption, but it could be driven by learning (followed by forgetting),
network investments (e.g. connections with friends ebb and flow if maintained or neglected), or more
nuanced habit formation mechanisms involving cues or automaticity (e.g. Laibson 2001; Bernheim and
Rangel 2004; Steiny Wellsjo 2021). Second, if so many of our participants perceive self-control problems
and use (and are willing to pay for) the Phone Dashboard time limit functionality, why isn’t there higher
demand for commercial digital self-control tools? Only 5 percent of our sample reported using any apps
to limit their smartphone use at baseline. Potential explanations include that our experimental setting or
selected set of participants overstates demand for commitment, that commercial self-control tools are too
expensive or are ineffective because it’s too easy to evade them or substitute across devices, that people
aren’t aware of existing tools, that the time misallocated due to temptation is not very valuable, or that the

29
commitment and flexibility features we built into Phone Dashboard were better suited to people’s needs. We
leave these questions for future work.

References
Acland, Dan and Vinci Chow. 2018. “Self-Control and Demand for Commitment in Online Game Playing:
Evidence from a Field Experiment.” Journal of the Economic Science Association 4 (1):46–62.

Acland, Dan and Matthew R. Levy. 2012. “Naivete, Projection Bias, and Habit Formation in Gym Atten-
dance.” Working Paper: GSPP13-002.

———. 2015. “Naiveté, Projection Bias, and Habit Formation in Gym Attendance.” Management Science
61 (1):146–160.

Allcott, Hunt, Luca Braghieri, Sarah Eichmeyer, and Matthew Gentzkow. 2020. “The Welfare Effects of
Social Media.” American Economic Review 110 (3):629–76.

Allcott, Hunt, Matthew Gentzkow, and Lena Song. 2020. “Digital Addiction.” AEA RCT Registry. Septem-
ber 27. 10.1257/rct.5796-2.0.

Allcott, Hunt, Joshua Kim, Dmitry Taubinsky, and Jonathan Zinman. 2021. “Are High-Interest Loans Preda-
tory? Theory And Evidence From Payday Lending.” Review of Economic Studies, forthcoming.

Allcott, Hunt and Todd Rogers. 2014. “The Short-Run and Long-Run Effects of Behavioral Interventions:
Experimental Evidence from Energy Conservation.” American Economic Review 104 (10):3003–37.

Alter, Adam. 2018. Irresistible: the Rise of Addictive Technology and the Business of Keeping Us Hooked.
Penguin Press.

Anderson, Michael L. 2008. “Multiple Inference and Gender Differences in the Effects of Early Interven-
tion: A Reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects.” Journal of the
American Statistical Association 103 (484):1481–1495.

Andreassen, Cecilie Schou, Torbjørn Torsheim, Geir Scott Brunborg, and Ståle Pallesen. 2012. “Develop-
ment of a Facebook Addiction Scale.” Psychological Reports 110 (2):501–517.

Andreoni, James and Charles Sprenger. 2012a. “Estimating Time Preferences from Convex Budgets.” Amer-
ican Economic Review 102 (7):3333–3356.

———. 2012b. “Risk Preferences Are Not Time Preferences.” American Economic Review 102 (7):3357–
3376.

Ashraf, Nava, Dean Karlan, and Wesley Yin. 2006. “Tying Odysseus to the Mast: Evidence from a Com-
mitment Savings Product in the Philippines.” The Quarterly Journal of Economics 121 (2):673–697.

Augenblick, Ned. 2018. “Short-Term Discounting of Unpleasant Tasks.” Working Paper.

30
Augenblick, Ned, Muriel Niederle, and Charles Sprenger. 2015. “Working Over Time: Dynamic Inconsis-
tency In Real Effort Tasks.” The Quarterly Journal of Economics 130 (3):1067–1115.

Augenblick, Ned and Matthew Rabin. 2019. “An Experiment on Time Preference and Misprediction in
Unpleasant Tasks.” The Review of Economic Studies 86 (3):941–975.

Auld, M Christopher and Paul Grootendorst. 2004. “An Empirical Analysis of Milk Addiction.” Journal of
Health Economics 23 (6):1117–1133.

Bai, Liag, Benjamin Handel, Edward Miguel, and Gautam Rao. 2018. “Self-Control and Demand for
Preventive Health: Evidence from Hypertension in India.” NBER Working Paper No. 23727.

Banerjee, Abhijit and Sendhil Mullainathan. 2010. “The Shape of Temptation: Implications for the Eco-
nomic Lives of the Poor.” NBER Working Paper No. 15973.

Becker, Gary S, Michael Grossman, and Kevin M Murphy. 1994. “An Empirical Analysis of Cigarette
Addiction.” The American Economic Review 84 (3):396–418.

Becker, Gary S and Kevin M Murphy. 1988. “A Theory of Rational Addiction.” Journal of Political Economy
96 (4):675–700.

Benjamini, Yoav and Yosef Hochberg. 1995. “Controlling the False Discovery Rate: A Practical and Pow-
erful Approach to Multiple Testing.” Journal of the Royal Statistical Society. Series B (Methodological)
57 (1):289–300.

Bernedo, Maria, Paul J Ferraro, and Michael Price. 2014. “The Persistent Impacts of Norm-Based Messag-
ing and Their Implications for Water Conservation.” Journal of Consumer Policy 37 (3):437–452.

Bernheim, B. Douglas and Antonio Rangel. 2004. “Addiction and Cue-Triggered Decision Processes.”
American Economic Review 94 (5):1558.

Beshears, John, James J. Choi, Christopher Harris, David Laibson, Brigitte C Madrian, and Jung Sakong.
2015. “Self-Control and Commitment: Can Decreasing the Liquidity of a Savings Account Increase
Deposits?” NBER Working Paper No. 21474.

Beshears, John and Katherine Milkman. 2017. “Creating Exercise Habits Using Incen-
tives: The Tradeoff between Flexibility and Routinization.” Working Paper, available at
https://www.semanticscholar.org/paper/Creating-Exercise-Habits-Using-Incentives-

Bianchi, Adriana and James G Phillips. 2005. “Psychological Predictors of Problem Mobile Phone Use.”
CyberPsychology & Behavior 8 (1):39–51.

Brandon, Alec, Paul J Ferraro, John A List, Robert D Metcalfe, Michael K Price, and Florian Rundham-
mer. 2017. “Do The Effects of Social Nudges Persist? Theory and Evidence from 38 Natural Field
Experiments.” NBER Working Paper No. 23277.

Brown, Eileen. 2019. “Americans spend far more time on their smartphones than they think.” Available at
https://www.zdnet.com/article/americans-spend-far-more-time-on-their-smartphones-than-they-think/.

31
Bursztyn, Leonardo, Davide Cantoni, David Y Yang, Noam Yuchtman, and Y Jane Zhang. 2020. “Persistent
Political Engagement: Social Interactions and the Dynamics of Protest Movements.” NBER Conference
Paper.

Busse, Meghan R, Devin G Pope, Jaren C Pope, and Jorge Silva-Risso. 2015. “The Psychological Effect of
Weather on Car Purchases.” Quarterly Journal of Economics 130 (1):371–414.

Carrera, Mariana, Heather Royer, Mark Stehr, and Justin Sydnor. 2018. “Can financial incentives help
people trying to establish new habits? Experimental evidence with new gym members.” Journal of
Health Economics 58:202–214.

Carrera, Mariana, Heather Royer, Mark Stehr, Justin Sydnor, and Dmitry Taubinsky. 2021. “Who Chooses
Commitment? Evidence and Welfare Implications.” Working Paper.

Carroll, Gabriel D, James J Choi, David Laibson, Brigitte C Madrian, and Andrew Metrick. 2009. “Optimal
Defaults and Active Decisions.” The Quarterly Journal of Economics 124 (4):1639–1674.

Casaburi, Lorenzo and Rocco Macchiavello. 2019. “Demand and Supply of Infrequent Payments as a
Commitment Device: Evidence from Kenya.” American Economic Review 109 (2):523–555.

Chaloupka, Frank. 1991. “Rational Addictive Behavior and Cigarette Smoking.” Journal of Political Econ-
omy 99 (4):722–742.

Chaloupka, Frank and Kenneth Warner. 1999. “The Economics of Smoking.” NBER Working Paper No.
7047.

Chaloupka, Frank J, Matthew R Levy, and Justin S White. 2019. “Estimating Biases in Smoking Cessation:
Evidence from a Field Experiment.” NBER Working Paper No. 26522.

Charness, Gary and Uri Gneezy. 2009. “Incentives to Exercise.” Econometrica 77 (3):909–931.

Collis, Avinash and Felix Eggers. 2019. “Effects of Restricting Social Media Usage.” Available at SSRN:
https://ssrn.com/abstract=3518744.

DellaVigna, Stefano and Ulrike Malmendier. 2006. “Paying Not to Go to the Gym.” American Economic
Review 96 (3):694–719.

Deloitte. 2018. “2018 Global Mobile Consumer Survey: US Edition.” Avail-

able at https://www2.deloitte.com/content/dam/Deloitte/us/Documents/technology-media-
telecommunications/us-tmt-global-mobile-consumer-survey-exec-summary-2018.pdf.

Do, Quy Toan and Hanan G Jacoby. 2020. “Sophisticated Policy with Naive Agents: Habit Formation and
Piped Water in Vietnam.” Available at SSRN: https://ssrn.com/abstract=3571024.

Duflo, Esther, Michael Kremer, and Jonathan Robinson. 2011. “Nudging Farmers to Use Fertilizer: Theory
and Experimental Evidence from Kenya.” American Economic Review 101 (6):2350–2390.

Ericson, Keith Marzilli and David Laibson. 2019. Intertemporal Choice, vol. 2, chap. 1. Elsevier, 1 ed.

Exley, Christine L. and Jeffrey K. Naecker. 2017. “Observability Increases the Demand for Commitment
Devices.” Management Science 63 (10):3147–3529.

32
Eyal, Nir. 2020. Indistractable: How to Control Your Attention and Choose Your Life. Bloomsbury Publish-
ing PLC.

Fang, Hanming and Dan Silverman. 2004. “Time Inconsistency and Welfare Program Participation: Evi-
dence from the NLSY.” Cowles Foundation Discussion Paper No. 1465.

Ferraro, Paul J, Juan Jose Miranda, and Michael K Price. 2011. “The Persistence of Treatment Effects
with Norm-Based Policy Instruments: Evidence from a Randomized Environmental Policy Experiment.”
American Economic Review 101 (3):318–22.

Fujiwara, Thomas, Kyle Meng, and Tom Vogl. 2016. “Habit Formation in Voting: Evidence from Rainy
Elections.” American Economic Journal: Applied Economics 8 (4):160–188.

Gerber, Alan S, Donald P Green, and Ron Shachar. 2003. “Voting May be Habit-Forming: Evidence from
a Randomized Field Experiment.” American Journal of Political Science 47 (3):540–550.

Gine, Xavier, Dean Karlan, and Jonathan Zinman. 2010. “Put Your Money Where Your Butt Is: A Commit-
ment Contract for Smoking Cessation.” American Economic Journal: Applied Economics 2:213–235.

Goda, Gopi Shah, Matthew R. Levy, Colleen Flaherty Manchester, Aaron Sojourner, and Joshua Tasoff.
2015. “The Role of Time Preferences and Exponential-Growth Bias in Retirement Savings.” NBER
Working Paper No. 21482.

Gosnell, Greer K, John A List, and Robert D Metcalfe. 2020. “The Impact of Management Practices
on Employee Productivity: A Field Experiment With Airline Captains.” Journal of Political Economy
128 (4):1195–1233.

Griffiths, Mark. 2005. “A “Components” Model of Addiction Within a Biopsychosocial Framework.” Jour-
nal of Substance Use 10 (4):191–197.

Gruber, Jonathan and Botond Köszegi. 2001. “Is Addiction “Rational”? Theory and Evidence.” Quarterly
Journal of Economics 116 (4):1261–1303.

Gul, Faruk and Wolfgang Pesendorfer. 2007. “Welfare without Happiness.” American Economic Review
97 (2):471–476.

Hawley, Josh. 2019. “S. 2314 (116th): SMART Act.” Available at

https://www.govtrack.us/congress/bills/116/s2314/text.

Hoong, Ruru. 2021. “Self Control and Smartphone Use: An Experimental Study of Soft Commitment
Devices.” Harvard Working Paper.

Hunt, Melissa G, Rachel Marx, Courtney Lipson, and Jordyn Young. 2018. “No More FOMO: Limit-
ing Social Media Decreases Loneliness and Depression.” Journal of Social and Clinical Psychology
37 (10):751–768.

Hussam, Reshmaan, Atonu Rabbani, Giovanni Reggiani, and Natalia Rigol. 2019. “Rational
Habit Formation: Experimental Evidence from Handwashing in India.” Available at SSRN:
https://ssrn.com/abstract=3040729.

33
Irvine, Mark. 2018. “Facebook Ad Benchmarks for YOUR Industry.”
https://www.wordstream.com/blog/ws/2017/02/28/facebook-advertising-benchmarks.
John, Anett. 2019. “When Commitment Fails - Evidence from a Field Experiment.” Management Science
66 (2):503–529.
John, Leslie K, George Loewenstein, Andrea B Troxel, Laurie Norton, Jennifer E Fassbender, and Kevin G
Volpp. 2011. “Financial Incentives for Extended Weight Loss: a Randomized, Controlled Trial.” Journal
of General Internal Medicine 26 (6):621–626.
Kaur, Supreet, Michael Kremer, and Sendhil Mullainathan. 2015. “Self-Control at Work.” Journal of
Political Economy 123 (6):1227–1277.
Kemp, Simon. 2020. “Digital 2020 Reports.” Available at wearesocial.com/digital-2020.
Kuchler, Theresa and Michaela Pagel. 2018. “Sticking to Your Plan: The Role of Present Bias for Credit
Card Paydown.” NBER Working Paper No. 24881.
Laibson, David. 1997. “Golden Eggs and Hyperbolic Discounting.” Quarterly Journal of Economics
112 (2):443–478.
———. 2001. “A Cue-Theory of Consumption.” The Quarterly Journal of Economics 116 (1):81–119.
———. 2018. “Private Paternalism, the Commitment Puzzle, and Model-Free Equilibrum.” AEA Papers
and Proceedings 108:1–21.
Laibson, David, Peter Maxted, Andrea Repetto, and Jeremy Tobacman. 2015. “Estimating Discount Func-
tions with Consumption Choices over the Lifecycle.” Working Paper.
Levitt, Steven D, John A List, and Sally Sadoff. 2016. “The Effect of Performance-Based Incentives on
Educational Achievement: Evidence from a Randomized Experiment.” NBER Working Paper No. 22107.
Liu, Zhuang, Michael Sockin, and Wei Xiong. 2020. “Data Privacy and Temptation.” NBER Working Paper
No. 27653.
Loewenstein, George, Ted O’Donoghue, and Matthew Rabin. 2003. “Projection Bias in Predicting Future
Utility.” Quarterly Journal of Economics 118 (4):1209–1248.
Madrian, Brigitte C and Dennis F Shea. 2001. “The Power of Suggestion: Inertia in 401(k) Participation
and Savings Behavior.” The Quarterly Journal of Economics 116 (4):1149–1187.
Makarov, Uliana. 2011. “Networking or not working: A model of social procrastination from communica-
tion.” Journal of Economic Behavior & Organization 2011 (80):574–585.
Marotta, Veronica and Alessandro Acquisti. 2017. “Online Distractions, Website Blockers, and Economic
Productivity: A Randomized Field Experiment.” Preliminary Draft.
Mosquera, Roberto, Mofioluwasademi Odunowo, Trent McNamara, Xiongfei Guo, and Ragan Petrie. 2019.
“The Economic Effects of Facebook.” Experimental Economics :1–28.
New York Post. 2017. “Americans Check Their Phones 80 Times a Day: Study.” Available at
https://nypost.com/2017/11/08/americans-check-their-phones-80-times-a-day-study/.

34
Newport, Cal. 2019. Digital Minimalism: Choosing a Focused Life in a Noisy World. Penguin Random
House.

O’Donoghue, Ted and Matthew Rabin. 1999. “Doing It Now or Later.” American Economic Review
89 (1):103–124. URL https://www.aeaweb.org/articles?id=10.1257/aer.89.1.103.

Paserman, M. Daniele. 2008. “Job Search and Hyperbolic Discounting: Structural Estimation and Policy
Evaluation.” The Economic Journal 118:1418–1452.

Pew Research Center. 2021. “Social Media Fact Sheet.” URL

https://www.pewresearch.org/internet/fact-sheet/social-media/.

Read, Danieal and Barbara Van Leeuwen. 1998. “Predicting Hunger: The Effects of Appetite and Delay on
Choice.” Organizational Behavior and Human Decision Processes 76 (2):189–205.

Rees-Jones, Alex and Kyle T. Rozema. 2020. “Price Isn’t Everything: Behavioral Response around Changes
in Sin Taxes.”

Royer, Heather, Mark Stehr, and Justin Sydnor. 2015. “Incentives, Commitments, and Habit Formation
in Exercise: Evidence from a Field Experiment with Workers at a Fortune-500 Company.” American
Economic Journal: Applied Economics 7 (3):51–84.

Sadoff, Sally, Anya Savikhin Samek, and Charles Sprenger. 2020. “Dynamic Inconsistency in Food Choice:
Experimental Evidence from a Food Desert.” Review of Economic Studies 87 (4):1954–1988.

Sagioglu, Christina and Tobias Greitemeyer. 2014. “Facebook’s Emotional Consequences: Why Facebook
Causes a Decrease in Mood and Why People Still Use It.” Computers in Human Behavior 35:359–363.

Schilbach, Frank. 2019. “Alcohol and Self-Control: A Field Experiment in India.” American Economic
Review 109 (4):1290–1322.

Shapiro, Jesse M. 2005. “Is There a Daily Discount Rate? Evidence from the Food Stamp Nutrition Cycle.”
Journal of Public Economics 89:303–325.

Shui, Haiyan and Lawrence M Ausubel. 2005. “Time Inconsistency in the Credit Card Market.” Working
Paper.

Skiba, Paige Marta and Jeremy Tobacman. 2018. “Payday Loans, Uncertainty, and Discounting: Explaining
Patterns of Borrowing, Repayment, and Default.” Working Paper.

Steiny Wellsjo, Alexandra. 2021. “Simple Actions, Complex Habits: Lessons from Hospital Hand Hygiene.”
Access on https://drive.google.com/file/d/1wbn6IuU0tMQ2VN6YHSWSCXv4v9pucKyK/view.

Strack, Philipp and Dmitry Taubinsky. 2021. “Dynamic Preference ”Reversals” and Time Inconsistency.”
NBER Working Paper No. 28961.

Toussaert, Severine. 2018. “Eliciting Temptation and Self-Control Through Menu Choices: A Lab Experi-
ment.” Econometrica 86 (3):859–889.

Tromholt, Morten. 2016. “The Facebook Experiment: Quitting Facebook Leads to Higher Levels of Well-
Being.” Cyberpsychology, Behavior, and Social Networking 19 (11):661–666.

35
United States Securities and Exchange Commission. 2020. “Official Facebook 2020 10-K report as filed
with SEC.” Available at https://d18rn0p25nwr6d.cloudfront.net/CIK-0001326801/4dd7fa7f-1a51-4ed9-
b9df-7f42cc3321eb.pdf.

Van Soest, Daan and Ben Vollaard. 2019. “Breaking Habits.” Working Paper.

Vanman, Eric J, Rosemary Baker, and Stephanie J Tobin. 2018. “The Burden of Online Friends: The Effects
of Giving Up Facebook on Stress and Well-Being.” The Journal of Social Psychology 158 (4):496–507.

Vox. 2020. “Tech Companies Tried to Help us Spend Less Time on Our Phones. It Didn’t Work.” Avail-
able at https://www.vox.com/recode/2020/1/6/21048116/tech-companies-time-well-spent-mobile-phone-
usage-data.

World Health Organization. 2018. “Gaming Disorder.” Available at

https://www.who.int/features/qa/gaming-disorder/en/.

Wurmser, Yoram. 2020. “US Mobile Time Spent 2020.” Available at https://www.emarketer.com/content/us-
mobile-time-spent-2020.

Zenith Media. 2019. “Consumers Will Spend 800 Hours Using Mobile Internet Devices This Year.” Avail-
able at https://www.zenithmedia.com/consumers-will-spend-800-hours-using-mobile-internet-devices-
this-year/.

36
 Table 1: Experiment Timeline and Sample Sizes

Phase Date Sample size

Recruitment and intake March 22 3,271,165 shown ads
- April 8 26,101 clicked on ads
18,589 passed screen
8,514 consented
5,320 finished intake survey
Survey 1 (baseline) April 12 4,134 began Survey 1
4,038 finished Survey 1
2,126 were randomized
Survey 2 May 3 2,068 began Survey 2
2,053 informed of treatment, of which:
2,048 were not in MPL group
2,032 finished Survey 2
Survey 3 May 24 1,993 began Survey 3
1,981 finished Survey 3
Survey 4 June 14 1,954 began Survey 4
1,948 finished Survey 4
Completion July 26 1,938 kept Phone Dashboard through July 26, of which:
1,933 were not in MPL group (“analysis sample”)

Table 2: Sample Demographics

(1) (2)
Analysis U.S.
sample adults
Income ($000s) 40.8 43.0
College 0.67 0.30
Male 0.39 0.49
White 0.72 0.74
Age 33.7 47.6
Period 1 phone use (minutes/day) 333.0 .
Period 1 FITSBY use (minutes/day) 152.8 .

Notes: Column 1 presents average demographics for our analysis sample, and column 2 presents average demographics
of American adults using data from the 2018 American Community Survey.

37
 Table 3: Empirical Moments for Restricted Model Estimation

(1) (2)
Point Confidence
Parameter Description estimate interval
τ3B Contemporaneous bonus effect (minutes/day) -55.9 [-61.7, -50.3]
τ4B Long-term bonus effect (minutes/day) -19.2 [-24.7, -13.7]
τ5B Long-term bonus effect (minutes/day) -12.3 [-18.1, -6.54]
τ2L Limit effect (minutes/day) -24.3 [-28.1, -20.4]
mC Control group misprediction (minutes/day) 6.13 [4.52, 7.72]
x̄1 Average baseline use (minutes/day) 153 [149, 157]
Notes: This table presents point estimates and bootstrapped 95 percent confidence intervals for the empirical moments
used for our primary estimates of the restricted model.

Table 4: Primary Parameter Estimates

(1) (2)
Restricted Unrestricted
model model
Parameter Description (units) (τ2B = 0, α = 1) (α = α̂)
λ Habit stock effect on consumption (unitless) 1.15 1.12
[0.609, 3.31] [0.572, 3.16]
ρ Habit formation (unitless) 0.299 0.302
[0.106, 0.493] [0.106, 0.498]

α Projection bias (unitless) 1 0.897

[0.584, 1.00]
η Price coefficient ($-day/hour2 ) −2.68 −2.75
[−2.98, −2.43] [−3.04, −2.51]
ζ Habit stock effect on marginal utility ($-day/hour2 ) 3.08 3.01
[1.65, 8.97] [1.55, 8.57]

γ − γ̃ Naivete about temptation ($/hour) 0.274 0.278

[0.201, 0.349] [0.205, 0.354]
γ Temptation ($/hour) 1.09 1.11
[0.884, 1.30] [0.903, 1.33]

κ̄ Average intercept ($/hour) −2.41 −2.24

[−3.62, −1.10] [−3.53, −0.803]
Notes: This table presents point estimates and bootstrapped 95 percent confidence intervals from the estimation strat-
egy described in Section 6.2 and Appendix E.3.

38
 Figure 1: Online and Offline Temptation
.8
(share “too much” – share “too little”)
.6
absolute value of

.4
.2
0
)

ia

od

ne

TV

es

l
ho
ai
(-1

-1

(-1

(-1

(-1

or
ed

fo

em

m
ho
t(

co
W
h
e

s
m

ga
en

c
y

tte
is

tp

ee

al
at
lth

ck
rc

al

ne
em

ar

re

k
Sl
a
e

he

rin
ci

de
sm

ga
he
Ex

tir

e
so

D
C

lin
vi
re

ci
un

se
se

ay

on

e
r

t
fo

ok
U
Ea
w

Pl

d
ve

Sm
ro

ea
Sa

← more perceived self-control problems | less perceived self-control problems →

Notes: This figure presents responses to the following question, which we asked participants in our experiment during
the baseline survey, “For each of the activities below, please tell us whether you think you do it too little, too much, or
the right amount.” The bars are ordered from left to right in order of largest to smallest absolute value of (share “too
little” – share “too much”).

39
 Figure 2: Experimental Design

Recruitment and intake survey

March 22-April 8

Survey 1 (baseline)
April 12

Period 1 (baseline)
Survey 2
May 3

Text message surveys (3 per week)

Bonus Bonus Control MPL
(25%) (75%) (0.2%)
Limit
(60%)
Limit Control
(40%)

Period 2

Survey 3
May 24

Period 3

Survey 4
June 14

Periods 4, 5

Study ends July 26

40
 Figure 3: Baseline Qualitative Evidence of Self-Control Problems

.4
.4

Fraction of sample
Fraction of sample

.3
.3

.2
.2

.1
.1

0
0

Not at all Slightly Moderately Very -100 -50 0 50 100

Interest in limits Ideal use change (percent)

.25
Fear missing out online
Wake up, check phone immediately
Use longer than intended

.2
Tell yourself just a few more minutes
Use to distract from personal issues Fraction of sample
Use to distract from anxiety/etc.
Use to relax to go to sleep .15
Try and fail to reduce use
Others are concerned about use
qname

.1

Feel anxious without phone

Difficult to put down phone
Annoyed at interruption in use
.05

Harms school/work performance

Lose sleep from use
Prefer phone to human interaction
Procrastinate by using phone
0

-5 0 5
0 .2 .4 .6 .8
Share of people who "often" or "always" Phone use makes life worse (left) or better (right)

Notes: This figure presents the distributions of four measures of smartphone addiction from the baseline survey.
Interest in limits is the answer to, “How interested are you to set limits on your phone use?” Ideal use change is the
answer to, “Relative to your actual use over the past 3 weeks, by how much would you ideally have [reduced/increased]
your screen time?” The bottom left panel presents the share of participants who responded “often” or “always” to each
of 16 questions modified from the Mobile Phone Problem Use Scale and the Bergen Facebook Addiction Scale. Phone
use makes life worse or better is the answer to, “To what extent do you think your smartphone use made your life better
or worse over the past 3 weeks?”

41
 Figure 4: Treatment Effects on FITSBY Use

0
●
Treatment effect (minutes/day)

●
−20 ●
● ●
●

−40

−60

Period 2 Period 3 Period 4 Period 5

● Bonus ● Limit
Notes: This figure presents effects of the bonus and limit treatments on FITSBY use using equation (4). FITSBY use
refers to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube.

42
 Figure 5: Effects on Smartphone Use by App
20
10
Treatment effect (minutes/day)

0
-10
-20
-30

Facebook Browser YouTube Instagram Snapchat Twitter Other

Bonus Limit

Notes: This figure presents effects of the bonus and limit treatments on smartphone use by app using equation (4).
The bonus effects are measured in period 3, while the limit effects are measured in periods 2–5. FITSBY use refers
to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube. FITSBY apps are in order of
decreasing period 1 use.

43
 Figure 6: Predicted vs. Actual FITSBY Use in Control Conditions
150
145
Usage (minutes/day)

140
135
130
125
120

Period 2 Period 3 Period 4 Period 5

Actual Survey t prediction

Survey t-1 prediction Survey t-2 prediction

Notes: This figure presents average actual FITSBY use by period and average predicted FITSBY use for that period, for
participants in the intersection of the Bonus Control and Limit Control groups. Period t is the three weeks immediately
after survey t, so “survey t prediction” is the prediction for period t made just prior to period t. FITSBY use refers to
screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube.

44
 Figure 7: Predicted vs. Actual Habit Formation

0
Treatment effect (minutes/day)

-20
-40
-60

Period 3 Period 4 Period 5

Actual Survey 2 MPL prediction

Survey 3 prediction

Notes: This figure presents the treatment effects of the bonus on FITSBY use and on predicted FITSBY use from
survey 3 using equation (4), as well as the average predicted bonus treatment effect elicited on survey 2 before the
bonus multiple price list. FITSBY use refers to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and
YouTube.

45
 Figure 8: Effects of Limits and Bonus on Survey Outcome Variables

Ideal use change

Addiction scale x (-1)

SMS addiction scale x (-1)

Phone makes life better

Subjective well-being

Survey index

-.2 0 .2 .4 .6

Treatment effect (standard deviations)

Bonus Limit

Notes: This figure presents effects of the bonus and limit treatments on survey outcome variables using equation
(4). The bonus effect is measured on survey 4, while the limit effect is measured on both surveys 3 and 4. Ideal
use change is the answer to, “Relative to your actual use over the past 3 weeks, by how much would you ideally
have [reduced/increased] your screen time?” Addiction scale is answers to a battery of 16 questions modified from
the Mobile Phone Problem Use Scale and the Bergen Facebook Addiction Scale. SMS addiction scale is answers to
shortened versions of the addiction scale questions delivered via text message. Phone makes life better is the answer to,
“To what extent do you think your smartphone use made your life better or worse over the past 3 weeks?” Subjective
well-being is answers to seven questions reflecting happiness, life satisfaction, anxiety, depression, concentration,
distraction, and sleep quality; anxiety, depression, and distraction are re-oriented so that more positive reflects better
subjective well-being. Survey index combines the previous five variables, weighting by the inverse of their covariance
at baseline.

46
 Figure 9: Model Identification

𝑥𝑥 ∗ (𝑠𝑠, 𝛾𝛾,
� 𝒑𝒑): 𝑥𝑥 ∗ (𝑠𝑠, 𝛾𝛾, 𝒑𝒑):
𝑥𝑥 ∗ (𝑠𝑠, 0, 𝒑𝒑): predicted actual
desired demand demand
demand 𝜏𝜏3𝐵𝐵 :
contemporaneous
bonus effect
B
𝑝𝑝 = 𝑝𝑝𝐵𝐵 J
A K
H
𝑚𝑚𝐶𝐶 : mis-
prediction
D G L
𝑝𝑝 = 0 C
Quantity
F 𝜏𝜏 𝐿𝐿 : temptation
effect
Price

47
 Figure 10: Effects of Temptation and Habit Formation on FITSBY Use

200

● ● ●
FITSBY use (minutes/day)

150 ●

● ●
100

●
●

●
50 ●

Baseline No naivete No temptation No habit No temptation

formation or habit formation

● Restricted model (α = 1) ● ^)
Unrestricted model (α = α

Notes: This figure presents point estimates and bootstrapped 95 percent confidence intervals for predicted steady-state
FITSBY use with different parameter assumptions, using equation (15).

48
Online Appendix Allcott, Gentzkow, and Song

Online Appendix

Digital Addiction
Hunt Allcott, Matthew Gentzkow, and Lena Song

49
Online Appendix Allcott, Gentzkow, and Song

Table of Contents
A Experimental Design Appendix 51
A.1 Variable Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

B Data Appendix 55

C Differences Between 2019 and the Study Period 59

D Model-Free Results Appendix 62

D.1 Validation of Predicted Use and Multiple Price List Responses . . . . . . . . . . . . . . . 68
D.2 Additional Estimates of Effects on Survey Outcome Variables . . . . . . . . . . . . . . . 77
D.3 Heterogeneous Treatment Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
D.4 Local Average Treatment Effects on Survey Outcomes . . . . . . . . . . . . . . . . . . . 83

E Unrestricted Model and Alternative Temptation Estimates 91

E.1 Key Theoretical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
E.2 Modeling the Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
E.3 Estimating Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
E.4 Empirical Moments and Estimation Details . . . . . . . . . . . . . . . . . . . . . . . . . 97
E.5 Alternative Temptation Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
E.6 Model Estimates with Sample Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

F Proofs of Propositions in Appendix E.1 106

F.1 Proof of Proposition 1: Euler Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
F.2 Proof of Proposition 2: Linear Policy Functions . . . . . . . . . . . . . . . . . . . . . . . 108
F.3 Proof of Lemma 1: Steady-State Convergence . . . . . . . . . . . . . . . . . . . . . . . . 116
F.4 Proof of Proposition 3: Steady-State Consumption . . . . . . . . . . . . . . . . . . . . . 116

G Derivations of Estimating Equations in Appendix E.3 117

G.1 Habit Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
G.2 Perceived Habit Formation, Price Response, and Habit Stock Effect on Marginal Utility . . 119
G.3 Naivete about Temptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
G.4 Temptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
G.5 Temptation with Multiple Goods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
G.6 Intercept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

H Counterfactual Simulations Appendix 129

50
Online Appendix Allcott, Gentzkow, and Song

A Experimental Design Appendix

Figure A1: Facebook Recruitment Ads

Notes: The ads at left and right were shown to users aged 18–34 and 35–64, respectively.

51
Online Appendix Allcott, Gentzkow, and Song

Figure A2: Phone Dashboard Screenshots

Snooze App Limit

Facebook You may resume using facebook after
Facebook 5:29 PM.
You have reached your limit of 1
Your daily limit for this app will expire in minute. You may resume using it in X
less than one minute of additional usage . How many additional minutes would
minutes if you snooze the app limit. you like?
CONTINUE You can revise your app limits in Phone
Dashboard, effective tomorrow. Additional Usage (Minutes)

SNOOZE APP LIMIT CONTINUE

CANCEL SNOOZE LIMIT

Notes: This figure presents screenshots of the Phone Dashboard app. The top left presents the day’s total usage by
app. The top middle shows how a user can set daily a daily usage limit for each app, effective tomorrow. The top right
shows the usage limits set for each app. The bottom left shows the warning users receive when they are within five
minutes or one minute of their limit. The bottom middle shows the message users receive when they reach the limit.
Users with the snooze functionality can resume using an app after a delay of X ∈ {0, 2, 5, 20} minutes. The bottom
right shows the option for a user to choose how many additional minutes to add to the daily limit after the snooze
52
delay. All participants had the usage information in the top left panel, while only the Limit group had the time limit
functionalities in the other panels.
Online Appendix Allcott, Gentzkow, and Song

A.1 Variable Definitions

Ideal use change. Some people say they use their smartphone too much and ideally would use it less.
Other people are happy with their usage or would ideally use it more. How do you feel about your overall
smartphone use over the past 3 weeks?

• I used my smartphone too much.

• I used my smartphone the right amount.

• I used my smartphone too little.

Relative to your actual use over the past 3 weeks, by how much would you ideally have [if “too much”:
reduced. If “too little”: increased] your smartphone use? Please give a number in percent. %

Addiction scale. Over the past 3 weeks, how often have you. . .

• Been worried about missing out on things online when not checking your phone?

• Checked social media, text messages, or email immediately after waking up?

• Used your phone longer than intended?

• Found yourself saying “just a few more minutes” when using your phone?

• Used your phone to distract yourself from personal problems?

• Used your phone to distract yourself from feelings of guilt, anxiety, helplessness, or depression?

• Used your phone to relax in order to go to sleep?

• Tried to reduce your phone use without success?

• Experienced that people close to you are concerned about the amount of time you use your phone?

• Felt anxious when you don’t have your phone?

• Found it difficult to switch off or put down your phone?

• Been annoyed or bothered when people interrupt you while you use your phone?

• Felt your performance in school or at work suffers because of the amount of time you use your phone?

• Lost sleep due to using your phone late at night?

• Preferred to use your phone rather than interacting with your partner, friends, or family?

• Put off things you have to do by using your phone?

53
Online Appendix Allcott, Gentzkow, and Song

Never, Rarely, Sometimes, Often, Always

SMS addiction scale.

• In the past 24 hours, did you use your phone longer than intended?

• In the past 24 hours, did your performance at school or work suffer because of the amount of time you
used your phone?

• In the past 24 hours, did you feel like you had an easy time controlling your screen time?

• In the past 24 hours, did you use your phone mindlessly?

• In the past 24 hours, did you use your phone because you were feeling down?

• In the past 24 hours, did using your phone keep you from working on something you needed to do?

• In the past 24 hours, would you ideally have used your phone less?

• Last night, did you lose sleep because of using your phone late at night?

• When you woke up today, did you immediately check social media, text messages, or email?

Please text back your answer on a scale from 1 (not at all) to 10 (definitely).

Phone makes life better. To what extent do you think your smartphone use made your life better or
worse over the past 3 weeks?
11-point scale from -5 (Makes my life worse) to 0 (Neutral) to 5 (Makes my life better)

Subjective well-being. Please tell us the extent to which you agree or disagree with each of the follow-
ing statements. Over the past 3 weeks, ...

• . . . I was a happy person

• . . . I was satisfied with my life

• . . . I felt anxious

• . . . I felt depressed

• . . . I could concentrate on what I was doing

• . . . I was easily distracted

• . . . I slept well

7-point scale from strongly disagree to neutral to strongly agree

54
Online Appendix Allcott, Gentzkow, and Song

B Data Appendix

Table A1: Response Rates

(a) Limit

(1) (2) (3) (4) (5) (6) (7) (8)

All Snooze Snooze Snooze Snooze No F-test
Control limits 0 2 5 20 snooze p-value
Completed survey 3 0.97 0.96 0.96 0.98 0.96 0.97 0.95 0.51
Completed survey 4 0.95 0.94 0.95 0.95 0.94 0.95 0.93 0.81
Have period 2 usage 1.00 1.00 0.99 0.99 1.00 1.00 1.00 0.23
Have period 3 usage 0.99 0.98 0.99 0.98 0.99 0.98 0.98 0.68
Have period 4 usage 0.98 0.97 0.99 0.98 0.97 0.97 0.96 0.37
Have period 5 usage 0.97 0.96 0.97 0.96 0.96 0.97 0.95 0.70

(b) Bonus

(1) (2) (3)

t-test
Control Treatment p-value
Completed survey 3 0.97 0.96 0.74
Completed survey 4 0.95 0.95 0.64
Have period 2 usage 1.00 1.00 0.16
Have period 3 usage 0.98 0.98 0.95
Have period 4 usage 0.98 0.97 0.84
Have period 5 usage 0.96 0.96 0.85

Notes: Columns 1 and 2 of Panel (a) present present response rates for Limit and Limit Control groups. Columns 3–7
present response rates for each of the snooze delay conditions within the Limit group. Column 8 presents the p-value
of an F-test of differences between the Limit Control and the separate snooze delay conditions. Columns 1 and 2 of
Panel (b) present response rates for Bonus and Bonus Control groups. Column 3 presents the p-value of a t-test of
differences between the Bonus and Bonus Control groups.

55
Online Appendix Allcott, Gentzkow, and Song

Table A2: Covariate Balance

(a) Limit

(1) (2) t-test

Treatment Control p-value
Variable Mean/SD Mean/SD (1)-(2)
Income ($000s) 40.15 41.76 0.35
(36.22) (37.84)
College 0.67 0.67 0.72
(0.47) (0.47)
Male 0.38 0.40 0.51
(0.49) (0.49)
White 0.70 0.74 0.13
(0.46) (0.44)
Age 33.61 33.79 0.76
(12.33) (12.35)
Period 1 FITSBY use (minutes/day) 151.96 154.07 0.64
(92.00) (99.19)
N 1150 783
F-test of joint significance (p-value) 0.65
F-test, number of observations 1933

(b) Bonus

(1) (2) t-test

Treatment Control p-value
Variable Mean/SD Mean/SD (1)-(2)
Income ($000s) 41.26 40.65 0.76
(39.16) (36.11)
College 0.67 0.67 0.75
(0.47) (0.47)
Male 0.41 0.38 0.26
(0.49) (0.49)
White 0.71 0.72 0.61
(0.46) (0.45)
Age 33.53 33.73 0.76
(12.17) (12.40)
Period 1 FITSBY use (minutes/day) 151.24 153.34 0.67
(91.97) (95.94)
N 479 1454
F-test of joint significance (p-value) 0.94
F-test, number of observations 1933

Notes: Panels (a) and (b) present tests of covariate balance for the Limit and Bonus treatment and control groups.

56
Online Appendix Allcott, Gentzkow, and Song

Figure A3: Most Popular Apps

60
1
.8

40
Share of users

Minutes/day
.4 .6

20
.2
0

0
ok

er

es

at

ic

s
ai
se

ap

ew
ub

in

on

us
ra

ch
ng

itt
Em
bo

am

am
ow

M
uT

Tw

Ph
ag

N
ap
se
ce

G
Br

re
Yo

st

Sn
es
Fa

In

St
M

Users at baseline Period 1 use

Notes: This figure presents the share of users that have each app and the average daily screen time in period 1 (base-
line). Period 1 use is across all users, not conditioning on whether or not they have the app.

57
Online Appendix Allcott, Gentzkow, and Song

Figure A4: Distribution of Baseline FITSBY Use

.15
Fraction of sample

.1
.05
0

0 200 400 600 800

Period 1 FITSBY use (minutes/day)

Notes: This figure presents a distribution of FITSBY use in period 1 (baseline). FITSBY use refers to screen time on
Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube.

Table A3: Descriptive Statistics for Survey Outcome Variables

Standard Minimum Maximum

Mean deviation value value
Ideal use change -19.0 21.4 -100 70
Addiction scale x (-1) -6.2 2.6 -16 0
SMS addiction scale x (-1) 1.7 3.1 -9 9
Phone makes life better 1.6 2.0 -5 5
Subjective well-being 0.2 2.5 -7 7

Notes: This table present descriptive statistics for the survey outcome variables at baseline.

58
Online Appendix Allcott, Gentzkow, and Song

C Differences Between 2019 and the Study Period

Figure A5: Effects of Coronavirus Outbreak on Free Time

Increased by more than 50%

Increased by 25-50%
Change in free time

Increased by 1-25%

No change

Decreased by 1-25%

Decreased by 25-50%

Decreased by more than 50%

0 .1 .2 .3

Fraction of sample

Notes: This figure presents the distribution of responses to the baseline survey question, “To what extent has the recent
coronavirus outbreak changed how much free time you have?”

59
Online Appendix Allcott, Gentzkow, and Song

Figure A6: Effects of Coronavirus on Smartphone Use

Increased phone usage

Effect of COVID-19 on phone use

No change

Decreased phone usage

Other

0 .2 .4 .6 .8

Fraction of sample

Notes: The baseline survey asked, “How has the recent coronavirus outbreak changed how you use your smartphone?”
We coded the responses as to whether they indicated increased, decreased, or unchanged smartphone use.

60
Online Appendix Allcott, Gentzkow, and Song

Figure A7: Self-Control Problems in 2019 versus Now

0
-10
Ideal use change

-20
-30

Difference in means = 4.844 (0.528)

-40

2019 Now
3
Phone use makes life better

2
1
0

Difference in means = -0.637 (0.052)

-1

2019 Now

Notes: This figure presents the mean (dots) and 25th and 75th percentiles (spikes) of responses to ideal use change
and phone use makes life better for 2019 and for the past 3 weeks, as reported on the baseline survey. Ideal use change
is the answer to, “Relative to your actual use [in 2019 / over the past 3 weeks], by how much would you ideally have
[reduced/increased] your screen time? Phone use makes life better is the answer to, “To what extent do you think your
smartphone use made your life better or worse [in 2019 / over the past 3 weeks]?”

61
Online Appendix Allcott, Gentzkow, and Song

D Model-Free Results Appendix

Figure A8: Ideal Use Change by App or Category

Facebook
Games
Instagram
Browser
YouTube
Streaming
Twitter
Messaging
Snapchat
Email
News
Maps
Phone
Music

-25 -20 -15 -10 -5 0 5

Ideal use change (percent)

Notes: This figure presents mean ideal use change by app or app category at baseline. Ideal use change is the answer
to, “Relative to your actual use over the past 3 weeks, by how much would you ideally have [reduced/increased] your
screen time?” We code “I don’t use this app at all” as 0, so these results reflect how much each app contributes to
overall temptation, not how tempting each app is for the subset of people who use it.

62
Online Appendix Allcott, Gentzkow, and Song

Figure A9: Effects of Bonus on FITSBY Use by Day for Periods 1 and 2

20
Treatment effect (minutes/day)

10
0
-10
-20

Period 1 Survey 2 Period 2

Notes: This figure presents differences in average FITSBY use between the Bonus and Bonus Control group for each
day of periods 1 and 2. The vertical line indicates the day of survey 2, when the bonus was announced. FITSBY use
refers to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube.

63
Online Appendix Allcott, Gentzkow, and Song

Figure A10: Effects of Bonus on FITSBY Use by Week

0 ●
●

●
●
●
Treatment effect (minutes/day)

● ●

−20 ●
●

−40

●
●
●

−60

4 5 6 7 8 9 10 11 12 13 14 15
Week of experiment
Notes: This figure presents effects of the bonus treatment on FITBSY use by week using equation (4). FITSBY use
refers to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube.

64
Online Appendix Allcott, Gentzkow, and Song

Figure A11: Distribution of User-Level Limit Tightness

.25
.2
Fraction of sample

.15
.1
.05
0

0 100 200 300 400

Periods 2 to 5 limit tightness (minutes/day)

Notes: This figure presents mean user-level limit tightness over periods 2–5. User-level limit tightness is the amount
by which a user’s limits would have hypothetically reduced overall screen time if applied to their baseline use without
snoozes; see equation (5).

65
Online Appendix Allcott, Gentzkow, and Song

Figure A12: Average Limit Tightness by App

25
20
Limit tightness (minutes/day)

15
10
5
0

Facebook Browser YouTube Instagram Snapchat Twitter Other

Notes: This figure presents average limit tightness by app over periods 2–5. Limit tightness is the amount by which
a user’s limits would have hypothetically reduced screen time if applied to their baseline use without snoozes; see
equation (5). FITSBY apps are in order of decreasing period 1 use.

66
Online Appendix Allcott, Gentzkow, and Song

Figure A13: Interaction Effects of Bonus and Limit by Period

25

0
Treatment effect (minutes/day)

●
●
●
−25 ●

−50

−75

Period 2 Period 3 Period 4 Period 5

Bonus ● Limit Limit x Bonus

Notes: This figure presents effects of bonus and limit treatments on FITSBY use using equation (4) with an additional
interaction term for participants in the intersection of the Limit and Bonus groups. FITSBY use refers to screen time
on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube.

67
Online Appendix Allcott, Gentzkow, and Song

Figure A14: Effects on Self-Reported FITSBY Use Change on Other Devices

10
5
Treatment effect (minutes/day)

0
-5
-10
-15

Bonus Limit

Notes: This figure presents the effects of bonus and limit treatments on self-reported change in FITSBY use on other
devices relative to the three weeks before the study using equation (4). FITSBY use refers to screen time on Facebook,
Instagram, Twitter, Snapchat, browsers, and YouTube. Self-reported changes are winsorized at 150 minutes.

D.1 Validation of Predicted Use and Multiple Price List Responses

Predicted use lines up well with actual use; see Appendix Figures A15 and A16. The $5 (instead of $1) pre-
diction accuracy reward slightly reduces the absolute value of the prediction error but has tightly estimated
zero effects on predicted use, actual use, and the level of the prediction error; see Appendix Table A4.
Multiple price lists are cognitively challenging, so we carry out several additional analyses to validate
that these valuations are informative about people’s preferences. First, participants’ valuations of the bonus
are correlated with the amount of money they could expect to earn; see Appendix Figure A19. Second, the
limit valuation and the behavior change premium (defined in Section E.3) are correlated with each other and
with limit tightness, ideal use change, addiction scale, SMS addiction scale, and other variables in expected
ways; see Appendix Table A5. Third, after the bonus MPL, we asked people to “select the statement that
best describes your thinking when trading off the Screen Time Bonus against the fixed payment.” 24 percent
responded that “I wanted to give myself an incentive to use my phone less over the next three weeks, even
though it might result in a smaller payment,” and this group had a substantially higher average behavior
change premium; see Appendix Figures A20 and A21.

68
Online Appendix Allcott, Gentzkow, and Song

Figure A15: Predicted vs. Actual FITSBY Use in Control

500

400
Actual FITSBY use (minutes/day)

Count
300
60

40

200 20

100

0 100 200 300 400 500

Predicted FITSBY use (minutes/day)
Notes: This figure presents the number of Control group participants in each cell of actual and predicted FITSBY use
across periods 2–4, using predictions from the survey just before each period. FITSBY use refers to screen time on
Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube.

69
Online Appendix Allcott, Gentzkow, and Song

Figure A16: Histogram of Actual Minus Predicted FITSBY Use in Control Group

0.20
Fraction of control group

0.15

0.10

0.05

0.00

−100 0 100
Actual minus predicted FITSBY use (minutes/day)
Notes: This figure presents the distribution of the difference between actual and predicted FITSBY use across periods
2–4 in the Control group, using predictions from the survey just before each period. FITSBY use refers to screen time
on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube.

Table A4: Effect of Prediction Accuracy Reward

(1) (2) (3) (4)

Absolute value of
Predicted Actual Predicted - predicted - actual
use use actual use use
High prediction reward 1.219 3.343 -2.207 -2.379
(2.582) (2.386) (1.691) (1.435)

Constant 118.9 116.7 2.300 35.22

(1.908) (1.670) (1.376) (1.212)

Notes: This table presents the effects of being offered the higher Prediction Reward ($5 instead of $1 for predicting
within 15 minutes of actual screen time) on predicted and actual FITSBY use in minutes per day. FITSBY use refers
to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube. Standard errors are in parentheses.

70
Online Appendix Allcott, Gentzkow, and Song

Figure A17: Valuation of Limit Functionality

.25
.2
Fraction of sample

.15
.1
.05
0

-10 0 10 20 30

Valuation of limit functionality ($)

Notes: This figure presents the distribution of valuations of access to the limit functionality for the next three weeks,
as elicited in a multiple price list on survey 3. Valuations above $20 are plotted at $25, and valuations below $-1 are
plotted at $-5.

71
Online Appendix Allcott, Gentzkow, and Song

Figure A18: Valuation of Screen Time Bonus

.15
Fraction of sample

.1
.05
0

0 50 100 150 200

Valuation of bonus ($)

Notes: This figure presents the distribution of valuations of the Screen Time Bonus incentive, as elicited on survey 2.
Valuations above $150 are plotted at $175.

72
Online Appendix Allcott, Gentzkow, and Song

Figure A19: Valuation of Bonus vs. Predicted Bonus Earnings

150
Predicted earnings from bonus ($)

Count
100 80

60

40

20
50 0

0 50 100 150
Valuation of bonus ($)
Notes: This figure presents the number of participants in each cell of predicted earnings from the Screen Time Bonus
(given the participant’s Bonus Benchmark and predicted FITSBY use) and valuation of the bonus, as elicited on survey
2.

73
Online Appendix Allcott, Gentzkow, and Song

Table A5: Correlations between Temptation and Addiction Measures

Behavior Valuation Limit Interest Ideal use Addiction SMS Phone

change of limit tightness in limits change scale addiction makes life
premium × (-1) scale better × (-1)
Behavior change
1 . . . . . . .
premium
Valuation
0.116 1 . . . . . .
of limit
Limit
0.471 0.199 1 . . . . .
tightness
Interest in
0.032 0.146 0.204 1 . . . .
limits
Ideal use
0.117 0.112 0.218 0.319 1 . . .
change x (-1)
Addiction
0.267 0.078 0.243 0.356 0.435 1 . .
scale
SMS addiction
0.272 0.132 0.259 0.312 0.345 0.651 1 .
scale
Phone makes
0.022 0.082 0.154 0.295 0.392 0.303 0.234 1
life better x (-1)
Note: The behavior change premium is the difference between the valuation of the Screen Time Bonus and the modeled
valuation if the consumer believed herself to be time consistent. Interest in limits, ideal use change, addiction scale,
SMS addiction scale, and phone makes life better are from survey 1.

74
Online Appendix Allcott, Gentzkow, and Song

Figure A20: Reported Reasoning on Screen Time Bonus Multiple Price List
.6
Fraction of sample

.4
.2
0

Only wanted Wanted incentive Don't want pressure Other

to maximize to use phone to use phone
earnings less less

Notes: After the bonus multiple price list, survey 2 asked participants to “select the statement that best describes
your thinking when trading off the Screen Time Bonus against the fixed payment.” This figure presents the share of
participants who selected each answer.

75
Online Appendix Allcott, Gentzkow, and Song

Figure A21: Behavior Change Premium by Reported Reasoning

30
Behavior change premium ($)

20
10
0
-10

Only wanted Wanted incentive Don't want pressure Other

to maximize to use phone to use phone
earnings less less

Notes: The behavior change premium is the difference between the valuation of the Screen Time Bonus and the
modeled valuation if the consumer believed herself to be time consistent. After the bonus multiple price list, survey 2
asked participants to “select the statement that best describes your thinking when trading off the Screen Time Bonus
against the fixed payment.” This figure presents means and 95 percent confidence intervals of the behavior change
premium by responses to that question.

76
Online Appendix Allcott, Gentzkow, and Song

D.2 Additional Estimates of Effects on Survey Outcome Variables

Figure A22: Effects of Limits and Bonus on Survey Outcomes on Surveys 3 and 4

Ideal use change

Addiction scale x (-1)

SMS addiction scale x (-1)

Phone makes life better

Subjective well-being

Survey index

-.4 -.2 0 .2 .4 .6

Treatment effect (standard deviations)

Bonus: Survey 3 Bonus: Survey 4 Limit: Survey 3 Limit: Survey 4

Notes: This figure presents effects of the bonus and limit treatment on survey outcome variables using equation (4),
allowing separate coefficients for effects on surveys 3 vs. 4. Ideal use change is the answer to, “Relative to your actual
use over the past 3 weeks, by how much would you ideally have [reduced/increased] your screen time?” Addiction
scale is answers to a battery of 16 questions modified from the Mobile Phone Problem Use Scale and the Bergen
Facebook Addiction Scale. SMS addiction scale is answers to shortened versions of the addiction scale questions
delivered via text message. Phone makes life better is the answer to, “To what extent do you think your smartphone use
made your life better or worse over the past 3 weeks?” Subjective well-being is answers to seven questions reflecting
happiness, life satisfaction, anxiety, depression, concentration, distraction, and sleep quality; anxiety, depression,
and distraction are re-oriented so that more positive reflects better subjective well-being. Survey index combines the
previous five variables, weighting by the inverse of their covariance at baseline.

77
Online Appendix Allcott, Gentzkow, and Song

Table A6: Treatment Effects

(a) Bonus
(1) (2) (3) (4) (5) (6)
Treatment Standard Treatment Standard P-value Sharpened
effect error effect error FDR-
(original (original (SD (SD adjusted
units) units) units) units) q-value
Ideal use change 9.0 1.6 0.41 0.074 0.000 0.000
Addiction scale x (-1) 0.44 0.10 0.16 0.037 0.000 0.000
SMS addiction scale x (-1) 0.42 0.13 0.14 0.041 0.001 0.004
Phone makes life better 0.042 0.090 0.021 0.045 0.64 0.78
Subjective well-being 0.23 0.10 0.090 0.040 0.026 0.09
Survey index 0.17 0.031 0.24 0.044 0.000 0.000

(b) Limit
(1) (2) (3) (4) (5) (6)
Treatment Standard Treatment Standard P-value Sharpened
effect error effect error FDR-
(original (original (SD (SD adjusted
units) units) units) units) q-value
Ideal use change 5.1 0.75 0.23 0.034 0.000 0.000
Addiction scale x (-1) 0.21 0.071 0.078 0.027 0.004 0.008
SMS addiction scale x (-1) 0.36 0.090 0.12 0.028 0.000 0.000
Phone makes life better 0.33 0.064 0.16 0.032 0.000 0.000
Subjective well-being 0.10 0.075 0.040 0.030 0.18 0.24
Survey index 0.13 0.020 0.18 0.029 0.000 0.000

Notes: This table presents effects of the bonus and limit treatments on survey outcome variables using equation
(4). The bonus effect is measured on survey 4, while the limit effect is measured on both surveys 3 and 4. Ideal
use change is the answer to, “Relative to your actual use over the past 3 weeks, by how much would you ideally
have [reduced/increased] your screen time?” Addiction scale is answers to a battery of 16 questions modified from
the Mobile Phone Problem Use Scale and the Bergen Facebook Addiction Scale. SMS addiction scale is answers to
shortened versions of the addiction scale questions delivered via text message. Phone makes life better is the answer to,
“To what extent do you think your smartphone use made your life better or worse over the past 3 weeks?” Subjective
well-being is answers to seven questions reflecting happiness, life satisfaction, anxiety, depression, concentration,
distraction, and sleep quality; anxiety, depression, and distraction are re-oriented so that more positive reflects better
subjective well-being. Survey index combines the previous five variables, weighting by the inverse of their covariance
at baseline. The effects in standard deviation units in column 3 match those reported on Figure 8.

78
Online Appendix Allcott, Gentzkow, and Song

Figure A23: Effects on Addiction Responses

Fear missing what happening online

Check social media/messages immediately after waking up

Use longer than intended

Tell yourself just a few more minutes

Use to distract from personal issues

Use to distract from anxiety/depression/etc.

Use to relax to go to sleep

Try and fail to reduce use

Others are concerned about use

Feel anxious without phone

Have difficulty putting down phone

Annoyed at interruption in use

Use harms school/work performance

Lose sleep from use

Prefer phone to human interaction

Procrastinate by using phone

-.06-.04-.02 0 .02 .04 .06

Treatment effect

Bonus Limit

Notes: This figure presents the effects of the bonus and limit treatments on individual items in the addiction scale
variable using equation (4). The bonus effect is measured on survey 4, while the limit effect is measured on both
surveys 3 and 4. The direction of the effects in this figure are opposite those in the main figures, because addiction
scale is multiplied by -1 in those figures.

79
Online Appendix Allcott, Gentzkow, and Song

Figure A24: Effects on SMS Addiction Responses

Use longer than intended

Use harms school/work performance

Easy to control screen time x (-1)

Use mindlessly

Use because felt down

Use kept from working on something needed

Ideally used phone less

Lose sleep from use

Check social media/messages immediately after waking up

-.2 -.15 -.1 -.05 0 .05 .1 .15 .2

Treatment effect

Bonus Limit

Notes: This figure presents the effects of the bonus and limit treatments on individual items in the SMS addiction
scale variable using equation (4). The bonus effect is measured on survey 4, while the limit effect is measured on
both surveys 3 and 4. The direction of the effects in this figure are opposite those in the main figures, because SMS
addiction scale is multiplied by -1 in those figures.

80
Online Appendix Allcott, Gentzkow, and Song

Figure A25: Effects on Subjective Well-Being Responses

Was happy

Was satisfied with life

Felt anxious x (-1)

Felt depressed x (-1)

Could concentrate

Was easily distracted x (-1)

Slept well

Happy <-> depressed index

Concentrate <-> sleep index

-.09 -.06 -.03 0 .03 .06 .09

Treatment effect

Bonus Limit

Notes: This figure presents the effects of the bonus and limit treatments on individual items in the subjective well-
being variable using equation (4). The bonus effect is measured on survey 4, while the limit effect is measured on both
surveys 3 and 4.

81
Online Appendix Allcott, Gentzkow, and Song

D.3 Heterogeneous Treatment Effects

Figure A26: Heterogeneous Effects of Limits and Bonus on Survey Index

(a) Bonus (b) Limit

Education Education

Age Age

Female Female

Baseline usage Baseline usage

Restriction index Restriction index

Addiction index Addiction index

0 .1 .2 .3 .4 .5 0 .1 .2 .3 .4

Treatment effect Treatment effect

(standard deviations) (standard deviations)

Above median Below median Above median Below median

Notes: This figure presents heterogeneous effects of the bonus and limit treatments on survey index, the inverse-
covariance weighted average of five measures of smartphone addiction and subjective well-being, using equation (4).
The bonus effect is measured on survey 4, while the limit effect is measured on both surveys 3 and 4. Above-median
education includes people with a college degree or more, above-median age includes people 30 and older, and median
baseline FITSBY use is 137 minutes per day. Restriction index is a combination of interest in limits and ideal use
change. Addiction index is a combination of addiction scale and phone makes life better.

82
Online Appendix Allcott, Gentzkow, and Song

Figure A27: Heterogeneous Effects of Limits and Bonus on FITSBY Use

(a) Bonus (b) Limit

Education Education

Age Age

Female Female

Baseline usage Baseline usage

Restriction index Restriction index

Addiction index Addiction index

-80 -60 -40 -20 -30 -20 -10 0 10

Treatment effect (minutes/day) Treatment effect (minutes/day)

Above median Below median Above median Below median

Notes: This figure presents heterogeneous effects of the bonus and limit treatments on FITSBY use using equation (4).
The bonus effects are measured in period 3, while the limit effects are measured in periods 2–5. FITSBY use refers
to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube. Above-median education includes
people with a college degree or more, above-median age includes people 30 and older, and median baseline FITSBY
use is 137 minutes per day. Restriction index is a combination of interest in limits and ideal use change. Addiction
index is a combination of addiction scale and phone makes life better.

D.4 Local Average Treatment Effects on Survey Outcomes

Our pre-analysis plan specified that we would also estimate instrumental variables (IV) regressions with
previous period FITSBY use xi,t−1 as the endogenous variable:

Yit = τxi,t−1 + βt X i1 + νit + εi , (17)

instrumenting for xi,t−1 with Bi and Li interacted with t = 3 and t = 4 indicators. We combine data from
surveys 3 and 4 and let all coefficients other than τ vary across the two periods. Conceptually, this regression
combines the effects of the bonus and limit intervention, weighting the interventions by their effects on
FITSBY use. Because the limit treatment could affect survey outcomes through channels other than reduced
FITSBY use—for example, by giving people an increased feeling of control over their screen time—we do
not claim that the IV exclusion restriction necessarily holds.
Appendix Figure A28 presents local average treatment effects estimated using equation (17), combin-
ing effects from both treatments. Appendix Figures A29–A34 study heterogeneity along the six pre-specified

83
Online Appendix Allcott, Gentzkow, and Song

moderators. The results are qualitatively similar to Figures 8 and A26, except that the estimates are slightly
more precise, as would be expected from combining effects of two interventions. Note that since the aver-
age effects of both interventions are about the same for people with low versus high baseline use (Figure
A26), the local average treatment effects of reduced use are much larger for people with low baseline use
(Appendix Figure A32).

Figure A28: Local Average Treatment Effects of FITSBY Use on Survey Outcome Variables

Ideal use change

Addiction scale x (-1)

SMS addiction scale x (-1)

Phone makes life better

Subjective well-being

Survey index

0 .2 .4 .6

Treatment effect (standard deviations per hour/day of use)

Notes: This figure presents local average treatment effects of FITSBY use on survey outcome variables using equation
(17). We instrument for FITSBY use with Bonus and Limit group indicators interacted with period indicators. FITSBY
use refers to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube. Ideal use change is the
answer to, “Relative to your actual use over the past 3 weeks, by how much would you ideally have [reduced/increased]
your screen time?” Addiction scale is answers to a battery of 16 questions modified from the Mobile Phone Problem
Use Scale and the Bergen Facebook Addiction Scale. SMS addiction scale is answers to shortened versions of the
addiction scale questions delivered via text message. Phone makes life better is the answer to, “To what extent do you
think your smartphone use made your life better or worse over the past 3 weeks?” Subjective well-being is answers to
seven questions reflecting happiness, life satisfaction, anxiety, depression, concentration, distraction, and sleep quality;
anxiety, depression, and distraction are re-oriented so that more positive reflects better subjective well-being. Survey
index combines the previous five variables, weighting by the inverse of their covariance at baseline.

84
Online Appendix Allcott, Gentzkow, and Song

Figure A29: Heterogeneous Effects on Survey Outcome Variables by Education

Ideal use change

Addiction scale x (-1)

SMS addiction scale x (-1)

Phone makes life better

Subjective well-being

Survey index

-.2 0 .2 .4 .6

Treatment effect
(standard deviations per hour/day of use)

Above median education

Below median education

Notes: This figure presents local average treatment effects of FITSBY use on survey outcome variables using equation
(17), for above- and below-median education. We instrument for FITSBY use with Bonus and Limit group indicators
interacted with period indicators. FITSBY use refers to screen time on Facebook, Instagram, Twitter, Snapchat,
browsers, and YouTube. Ideal use change is the answer to, “Relative to your actual use over the past 3 weeks, by
how much would you ideally have [reduced/increased] your screen time?” Addiction scale is answers to a battery of
16 questions modified from the Mobile Phone Problem Use Scale and the Bergen Facebook Addiction Scale. SMS
addiction scale is answers to shortened versions of the addiction scale questions delivered via text message. Phone
makes life better is the answer to, “To what extent do you think your smartphone use made your life better or worse
over the past 3 weeks?” Subjective well-being is answers to seven questions reflecting happiness, life satisfaction,
anxiety, depression, concentration, distraction, and sleep quality; anxiety, depression, and distraction are re-oriented
so that more positive reflects better subjective well-being. Survey index combines the previous five variables, weighting
by the inverse of their covariance at baseline.

85
Online Appendix Allcott, Gentzkow, and Song

Figure A30: Heterogeneous Effects on Survey Outcome Variables by Age

Ideal use change

Addiction scale x (-1)

SMS addiction scale x (-1)

Phone makes life better

Subjective well-being

Survey index

0 .2 .4 .6 .8

Treatment effect
(standard deviations per hour/day of use)

Above median age

Below median age

Notes: This figure presents local average treatment effects of FITSBY use on survey outcome variables using equation
(17), for above- and below-median age. We instrument for FITSBY use with Bonus and Limit group indicators
interacted with period indicators. FITSBY use refers to screen time on Facebook, Instagram, Twitter, Snapchat,
browsers, and YouTube. Ideal use change is the answer to, “Relative to your actual use over the past 3 weeks, by
how much would you ideally have [reduced/increased] your screen time?” Addiction scale is answers to a battery of
16 questions modified from the Mobile Phone Problem Use Scale and the Bergen Facebook Addiction Scale. SMS
addiction scale is answers to shortened versions of the addiction scale questions delivered via text message. Phone
makes life better is the answer to, “To what extent do you think your smartphone use made your life better or worse
over the past 3 weeks?” Subjective well-being is answers to seven questions reflecting happiness, life satisfaction,
anxiety, depression, concentration, distraction, and sleep quality; anxiety, depression, and distraction are re-oriented
so that more positive reflects better subjective well-being. Survey index combines the previous five variables, weighting
by the inverse of their covariance at baseline.

86
Online Appendix Allcott, Gentzkow, and Song

Figure A31: Heterogeneous Effects on Survey Outcome Variables by Gender

Ideal use change

Addiction scale x (-1)

SMS addiction scale x (-1)

Phone makes life better

Subjective well-being

Survey index

0 .2 .4 .6 .8

Treatment effect
(standard deviations per hour/day of use)

Female
Male

Notes: This figure presents local average treatment effects of FITSBY use on survey outcome variables using equation
(17), for men versus women. We instrument for FITSBY use with Bonus and Limit group indicators interacted
with period indicators. FITSBY use refers to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and
YouTube. Ideal use change is the answer to, “Relative to your actual use over the past 3 weeks, by how much would
you ideally have [reduced/increased] your screen time?” Addiction scale is answers to a battery of 16 questions
modified from the Mobile Phone Problem Use Scale and the Bergen Facebook Addiction Scale. SMS addiction scale
is answers to shortened versions of the addiction scale questions delivered via text message. Phone makes life better
is the answer to, “To what extent do you think your smartphone use made your life better or worse over the past 3
weeks?” Subjective well-being is answers to seven questions reflecting happiness, life satisfaction, anxiety, depression,
concentration, distraction, and sleep quality; anxiety, depression, and distraction are re-oriented so that more positive
reflects better subjective well-being. Survey index combines the previous five variables, weighting by the inverse of
their covariance at baseline.

87
Online Appendix Allcott, Gentzkow, and Song

Figure A32: Heterogeneous Effects on Survey Outcome Variables by Baseline FITSBY Use

Ideal use change

Addiction scale x (-1)

SMS addiction scale x (-1)

Phone makes life better

Subjective well-being

Survey index

0 .2 .4 .6 .8 1

Treatment effect
(standard deviations per hour/day of use)

Above median period 1 usage

Below median period 1 usage

Notes: This figure presents local average treatment effects of FITSBY use on survey outcome variables using equation
(17), for above- and below-median baseline FITSBY use. We instrument for FITSBY use with Bonus and Limit
group indicators interacted with period indicators. FITSBY use refers to screen time on Facebook, Instagram, Twitter,
Snapchat, browsers, and YouTube. Ideal use change is the answer to, “Relative to your actual use over the past 3 weeks,
by how much would you ideally have [reduced/increased] your screen time?” Addiction scale is answers to a battery
of 16 questions modified from the Mobile Phone Problem Use Scale and the Bergen Facebook Addiction Scale. SMS
addiction scale is answers to shortened versions of the addiction scale questions delivered via text message. Phone
makes life better is the answer to, “To what extent do you think your smartphone use made your life better or worse
over the past 3 weeks?” Subjective well-being is answers to seven questions reflecting happiness, life satisfaction,
anxiety, depression, concentration, distraction, and sleep quality; anxiety, depression, and distraction are re-oriented
so that more positive reflects better subjective well-being. Survey index combines the previous five variables, weighting
by the inverse of their covariance at baseline.

88
Online Appendix Allcott, Gentzkow, and Song

Figure A33: Heterogeneous Effects on Survey Outcome Variables by Restriction Index

Ideal use change

Addiction scale x (-1)

SMS addiction scale x (-1)

Phone makes life better

Subjective well-being

Survey index

-.2 0 .2 .4 .6 .8

Treatment effect
(standard deviations per hour/day of use)

Above median restriction index

Below median restriction index

Notes: This figure presents local average treatment effects of FITSBY use on survey outcome variables using equation
(17), for above- and below-median values of restriction index, a combination of interest in limits and ideal use change.
We instrument for FITSBY use with Bonus and Limit group indicators interacted with period indicators. FITSBY
use refers to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube. Ideal use change is the
answer to, “Relative to your actual use over the past 3 weeks, by how much would you ideally have [reduced/increased]
your screen time?” Addiction scale is answers to a battery of 16 questions modified from the Mobile Phone Problem
Use Scale and the Bergen Facebook Addiction Scale. SMS addiction scale is answers to shortened versions of the
addiction scale questions delivered via text message. Phone makes life better is the answer to, “To what extent do you
think your smartphone use made your life better or worse over the past 3 weeks?” Subjective well-being is answers to
seven questions reflecting happiness, life satisfaction, anxiety, depression, concentration, distraction, and sleep quality;
anxiety, depression, and distraction are re-oriented so that more positive reflects better subjective well-being. Survey
index combines the previous five variables, weighting by the inverse of their covariance at baseline.

89
Online Appendix Allcott, Gentzkow, and Song

Figure A34: Heterogeneous Effects on Survey Outcome Variables by Addiction Index

Ideal use change

Addiction scale x (-1)

SMS addiction scale x (-1)

Phone makes life better

Subjective well-being

Survey index

-.2 0 .2 .4 .6 .8

Treatment effect
(standard deviations per hour/day of use)

Above median addiction index

Below median addiction index

Notes: This figure presents local average treatment effects of FITSBY use on survey outcome variables using equation
(17), for above- and below-median values of addiction index, a combination of addiction scale and phone makes
life better. We instrument for FITSBY use with Bonus and Limit group indicators interacted with period indicators.
FITSBY use refers to screen time on Facebook, Instagram, Twitter, Snapchat, browsers, and YouTube. Ideal use
change is the answer to, “Relative to your actual use over the past 3 weeks, by how much would you ideally have
[reduced/increased] your screen time?” Addiction scale is answers to a battery of 16 questions modified from the
Mobile Phone Problem Use Scale and the Bergen Facebook Addiction Scale. SMS addiction scale is answers to
shortened versions of the addiction scale questions delivered via text message. Phone makes life better is the answer to,
“To what extent do you think your smartphone use made your life better or worse over the past 3 weeks?” Subjective
well-being is answers to seven questions reflecting happiness, life satisfaction, anxiety, depression, concentration,
distraction, and sleep quality; anxiety, depression, and distraction are re-oriented so that more positive reflects better
subjective well-being. Survey index combines the previous five variables, weighting by the inverse of their covariance
at baseline.

90
Online Appendix Allcott, Gentzkow, and Song

E Unrestricted Model and Alternative Temptation Estimates

In this appendix, we estimate the unrestricted model and present alternative estimates of the temptation
parameter γ.

E.1 Key Theoretical Results

Three theoretical results are key to our estimation strategy: the Euler equation, linear policy functions, and
the steady state.
Euler equation. The first-order conditions of equation (2) for periods t and t +1 can be re-arranged into
an Euler equation characterizing the equilibrium relationship between consumption in periods t and t +1. To
simplify notation, define ut := ut (xt∗ ; st , pt ) as current utility, define x̃r := x̃r∗ (s̃r , γ̃, p r ) and ũr := ur (x̃r ; s̃r , pr )
∂ x̃r
as predicted consumption and utility for future periods r > t, and define λ̃r := ∂ s̃r as the predicted effect of
habit stock on consumption.

Proposition 1. Suppose ut (xt ; st , pt ) is given by equation (3) and (x0∗ , ..., xT∗ ) is a perception-perfect strategy
profile with differentiable strategies. Then for each t < T ,

 

ηxt∗ + ζ st + ξt − pt +γ = (1 − α)δ ρ η x̃t+1 + ζ s̃t+1 + ξt+1 − pt+1 +γ̃ + γ̃ λ̃t+1 − (ζ x̃t+1 + φ ) . (18)
 
| {z } | {z } | {z }
∂ ut /∂ xt ∂ ũt+1 /∂ x̃t+1 ∂ ũt+1 /∂ s̃t+1

Proof. See Appendix F.1.

With full myopia (δ = 0) or full projection bias (α = 1), consumers maximize current-period flow
utility, setting the left-hand side of equation (18) to zero. In a “rational” habit formation model with α = 0
and γ̃ = γ = 0, the right-hand side adds two effects. First, there is an adjacent complementarity effect where
people consume more in period t (driving down marginal utility ∂ ut /∂ xt ) if they expect to consume more in
t + 1 (i.e. if future marginal utility ∂ ũt+1 /∂ xt+1 is lower). Second, there is a direct habit stock effect where
people consume more in period t if the marginal utility from the resulting habit stock ∂ ũt+1 /∂ st+1 is higher.
Temptation adds two forces. First, the balance of the adjacent complementarity effect tilts toward in-
creased consumption, as γ is added to period t marginal utility and γ̃ is added to predicted period t + 1
marginal utility. Second, people reduce current consumption to avoid exacerbating perceived future over-
consumption, giving γ̃λt+1 on the right-hand side.
Linear policy functions. With quadratic flow utility, equilibrium consumption is linear in habit stock
with slope λt , and equilibrium predicted consumption is linear in habit stock with slope λ̃t . Furthermore, if

91
Online Appendix Allcott, Gentzkow, and Song

the consumer’s objective function is concave, λ and λ̃ are constant far from the time horizon. This argument
follows Gruber and Köszegi (2001).

Proposition 2. Suppose the conditions for Proposition 1 hold. Then for any t,

xt∗ (st , γ, pt ) = λt st + µt (γ) (19)

x̃t∗ (st , γ̃, pt ) = λ̃t st + µt (γ̃), (20)

where λt is a function of only {η, ζ , δ , ρ, α}, λ̃t is a function of only {η, ζ , δ , ρ}, and µt is linear in pt . Fur-
h λ̃t = λ̃for
thermore, if the objective function from equation (2) is concave, then limT →∞ λt = λ and limT →∞  i
any fixed t. Finally, limT →∞ µt = µ for any fixed t if pt and ξt are constant and −η > (1−α)δ ρ (ζ − η) 1 + ρ λ̃t+1 − ρζ .

Proof. See Appendix F.2. That appendix also provides an explicit condition that guarantees concavity.

Steady state. Over a period of time when strategies are well approximated by the limiting values λ
and µ, consumption converges to a steady state.

Lemma 1. Suppose that strategies in all periods take the form xt∗ (st , γ, pt ) = λ st + µ, where λ and µ are
constant. If ρ (1 + λ ) < 1, both xt∗ and st converge monotonically over time to steady-state values xss and
sss .

Proof. See Appendix F.3.

If consumption has reached a steady state, we can use the Euler equation to characterize its level in
closed form.

Proposition 3. Suppose that pt and ξt are constant and that consumption and habit stock are in steady state
with st = sss , xt = xss , and xss = ρ (sss + xss ). Then consumption can be written as
h   i
κ − (1 − (1 − α)δ ρ) p + (1 − α)δ ρ (ζ − η)mss − 1 + λ̃ γ̃ + γ
xss = 2 , (21)
−η − (1 − α)δ ρ(ζ − η) − ζ ρ−(1−α)δ
1−ρ
ρ

where κ := (1 − α)δ ρ(φ − ξ ) + ξ and mss := x̃t+1 − xss is steady-state misprediction.

Proof. See Appendix F.4.

The parameter restrictions required for Proposition 2 and Lemma 1 (including concavity) essentially
amount to requiring that perceived and actual habit formation are not too strong. We have confirmed that
these restrictions hold at the parameter estimates presented in Table 4.

92
Online Appendix Allcott, Gentzkow, and Song

E.2 Modeling the Experiment

We need additional notation to map the experiment’s treatments and data into the model and estimation.
We define xit to be participant i’s daily average FITSBY screen time during period t, x̃it to be participant i’s
predicted screen time elicited on a survey, and mit = xit − x̃it to be the difference between the two. The Bonus
and Bonus Control groups are denoted g ∈ {B, BC}, the Limit and Limit Control groups are g ∈ {L, LC}, and
the intersection of Bonus Control and Limit Control is g = C. We define ȳ := Ei yi as the expectatation over
participants of variable y, and yg := Ei∈g yi as the expectation over group g. τtg := xtg − xtgC and τ̃tg := x̃tg − x̃tgC
are the actual and predicted average treatment effects.
We model the Screen Time Bonus as a price pB = $2.50 per hour in period 3 plus a fixed payment FiB =
hours
$50 × ceil(xi1 day ), where ceil(·) rounds up to the nearest integer, giving participant i’s Bonus Benchmark.
In this appendix, we generalize the primary model from Section 6 by modeling the limit as an intervention
that eliminates share ω of temptation.
We define vBi as the valuation of the bonus elicited on survey 2, and we define vLi as the valuation of
access to the limit functionality elicited on survey 3. We assume that on survey t, consumers are aware of
period t projection bias when predicting period t consumption and are projection biased when determin-
ing their bonus and limit valuations. This assumption means that misprediction of period-t consumption is
driven only by naivete about temptation, and that bonus and limit valuations are driven only by perceived
temptation, not by an additional desire to offset projection bias. We acknowledge that alternative assump-
tions could be made.

E.3 Estimating Equations

Using the theoretical results from Appendix E.1, we can now derive equations that characterize how a con-
sumer from our unrestricted model would behave in our experiment. These equations parallel the equation
in Section 6.2, with additional terms that account for perceived habit formation. We assume that the dis-
count factor is δ = 0.997 per three-week period, consistent with a five percent annual discount rate. We
estimate the remaining parameters in stages, as described below. Appendix G presents formal derivations
and additional details.

Habit Formation

We first estimate λ and ρ from the decay of the bonus treatment effects. Even though λ is not a structural
parameter, it is easily identified and useful in estimating the other parameters. Using the habit stock evolution
formula and the linearity result in equation (19), we can write the period 4 bonus effect as the result of
decayed effects from periods 2 and 3: τ4B = λ ρτ3B + ρ 2 τ2B . Similarly, the period 5 effect results from the

cumulative decayed effects from periods 2–4: τ5B = λ ρτ4B + ρ 2 τ3B + ρ 3 τ2B . Rearranging gives a system of

two equations for λ and ρ:

93
Online Appendix Allcott, Gentzkow, and Song

τ4B
λ= (22)
ρτ3 + ρ 2 τ2B
B

τ5B
ρ= B . (23)
τ4 (1 + λ )

This non-linear system has two solutions when τ2B 6= 0, but in our data there is only one solution that satisfies
the requirement that ρ ≥ 0.
For estimation, we assume λ̃ = λ . This is reasonable because Figure 7 shows that participants predicted
the time path of bonus effects with reasonable accuracy, so calibrating equations (22) and (23) with predicted
τtB would not change the estimates much. To the extent that predictions differ from actual behavior, we prefer
to err on the side of using actual behavior instead of beliefs to estimate the model.

Perceived Habit Formation, Price Response, and Habit Stock Effect on Marginal Utility

After estimating λ and ρ, we estimate α, η, and ζ from the magnitude and decay of the bonus treatment
effects. For each of periods 2, 3, and 4, we difference the Euler equations for the Bonus and Bonus Control
groups and rearrange, giving a system of three equations for (1 − α), η, and ζ :

ητ2B
(1 − α) =  . (24)
δ ρ −pB + (η − ζ )τ̃3B + ζ ρτ2B
pB − ζ ρτ2B + (1 − α)δ ρ 2 ζ (1 − λ̃ ) ρτ2B + τ3B


η=
(25)
τ3B − (1 − α)δ ρ 2 λ̃ ρτ2B + τ3B
−ητ4B + (1 − α)δ ρ 2 η λ̃ ρ 2 τ2B + ρτ3B + τ4B


ζ=  
. (26)
ρτ3B + ρ 2 τ2B − (1 − α)δ ρ 2 1 − λ̃ ρ 2 τ2B + ρτ3B + τ4B

The first equation shows that as the anticipatory demand response in period 2 grows compared to the pre-
dicted demand response in period 3 (making τ2B /τ̃3B larger), we infer more perceived habit formation (smaller
α).

Naivete about Temptation

Next, we estimate naivete about temptation γ − γ̃ using the Control group’s difference between perceived
and actual consumption. To solve for γ − γ̃, we difference the actual versus perceived Euler equations for
group C, giving
h  i
γ − γ̃ = mCt · −η + (1 − α)δ ρ 2 (η − ζ )λ̃ + ζ . (27)

94
Online Appendix Allcott, Gentzkow, and Song

Temptation

We estimate temptation γ using three different strategies: the limit treatment effect and valuations of the
bonus and limit. Each strategy delivers an equation that we combine with equation (27) to form a system of
two equations for γ and γ̃.
Limit effect. Recall that we model the limit as an intervention that eliminates share ω of temptation,
starting in period 2. Thus, we can identify γ using an assumed ω plus the effect of the limit on consumption.
To solve for γ, we difference the Euler equations for periods 2 versus 3 for the Limit group compared to
Limit Control and rearrange, giving
h i
γ = ητ2L /ω − (1 − α)δ ρ (η − ζ )τ̃3L /ω + ζ ρτ2L /ω − γ̃ − γ̃ λ̃ . (28)

Our primary estimates in Section 6 use this equation, after setting ω = 1 and α = 1.
Bonus valuation. Since the bonus is like a commitment device that reduces future use, people with
perceived self-control problems will place higher value on the bonus. We can estimate perceived temptation
γ̃ from participants’ valuations. Our derivation follows Allcott, Kim, Taubinsky, and Zinman (2021), and
the approach also follows Acland and Levy (2012), Augenblick and Rabin (2019), Chaloupka, Levy, and
White (2019), and Carrera et al. (2021).
Let Vt (s̃t , ·) be the period t continuation value function conditional on s̃t , according to predicted con-
sumption and preferences before period t. This reflects preferences of a consumer filling out the multiple
price list on a survey before period t. Since utility is quasilinear in money, Vt (st , ·) is in units of period t
dollars.
The effect of a period 3 price increase from 0 to pB3 on the period 3 continuation value is

1
∆V3 pB := V3 s̃3 , p3 = pB3 −V3 (s̃3 , p3 = 0) = −pB3 · x̃3 (pB3 ) + x̃3 (0) − γ̃ · x̃3 (pB3 ) − x̃3 (0) ,





(29)
2

where x̃3 (p3 ) = x̃3∗ (s̃3 , γ̃, p 3 ) is shorthand for predicted period 3 consumption as a function of period 3
price. Figure 9 illustrates. The trapezoid ABCD is pB3 · 12 x̃3 (pB3 ) + x̃3 (0) : the survey taker’s prediction of

the consumer surplus loss from the price increase from the period 3 self’s perspective. The parallelogram
BCEF is −γ̃ · x̃3 (pB3 ) − x̃3 (0) : the predicted additional temptation reduction benefit from the survey taker’s

perspective.
The Screen Time Bonus combines a price change with a fixed payment of F B . Thus, the model predicts
that people filling out the bonus MPL would be indifferent between the bonus and a fixed payment of vB =
F B + ∆V3 (pB ). Taking the expectation over participants to allow mean-zero survey noise, substituting τ̃3B :=
Ei x̃i3 (pB3 ) − x̃i3 (0) and x̃¯3B+BC := Ei 1 x̃i3 (pB3 ) + x̃i3 (0) , and rearranging gives perceived temptation:
  

2

v̄B − F̄ B + pB3 x̃¯3B+BC

γ̃ = . (30)
−τ̃3B

95
Online Appendix Allcott, Gentzkow, and Song

The model predicts that if consumers perceive themselves to be time consistent (γ̃ = 0), the average bonus
valuation would equal the average valuation from the period 3 self’s perspective, F̄ B − pB3 x̃¯3B+BC . We refer
to the difference between the observed average valuation and the modeled time-consistent valuation (the
numerator of equation (30)) as “behavior change premium.” We infer more perceived temptation γ̃ from a
larger behavior change premium.
Limit valuation. People who perceive future temptation value the limit, as they perceive that it elimi-
nates share ω of temptation. We can estimate perceived temptation γ̃ using an assumed ω plus the valuation
the limit functionality. We solve for the modeled valuation similarly to how we solved for the bonus valua-
tion above.
The effect of a period 3 temptation reduction from γ̃ to (1 − ω)γ̃ on the period 3 continuation value is

2−ω
vL = V3 (s3 , γ̃3 = (1 − ω)γ̃) −V3 (s3 , γ̃3 = γ̃) = γ̃ · (x3∗ (γ̃) − x3∗ ((1 − ω)γ̃)) · , (31)
2

where x3∗ (γ̃3 ) is now shorthand for predicted period 3 consumption as a function of predicted period 3
temptation. Figure 9 illustrates. With ω = 1, the limit valuation is the deadweight loss reduction CEG from
the survey taker’s perspective from consuming the desired amount (x3∗ (0), point G) instead of the predicted
amount (x3∗ (γ̃), point C). The height of this triangle is γ̃ and the width is x3∗ (γ̃) − x3∗ (0), and thus the area is
γ̃ · (x3∗ (γ̃) − x3∗ (0)) · 21 . With ω < 1, the valuation vL equals the deadweight loss reduction trapezoid starting
to the right of point G and bounded by segment CE.
Taking the expectation over participants, substituting τ̃3L := Ei [x3∗ ((1 − ω)γ̃) − x3∗ (γ̃)], and rearranging
gives perceived temptation:

v̄L
γ̃ = . (32)
−τ̃3L (2 − ω)/2
We infer more perceived temptation γ̃ from higher valuation v̄L .

Intercept

Finally, we back out a heterogeneous intercept κi that explains observed consumption heterogeneity. Our
data do not allow us to separately identify φ (the direct effect of habit stock on utility) from ξ (the marginal
utility shifter), so κi includes both of these structural parameters. We assume that participant i’s observed
baseline consumption xi1 is in a steady state characterized by equation (21). Rearranging that equation gives

h   i
κi := (1 − α)δ ρ(φ − ξi ) + ξi = (1 − (1 − α)δ ρ) p − (1 − α)δ ρ (ζ − η) mss − 1 + λ̃ γ̃
ρ − (1 − α)δ ρ 2
 
− γ + xi1 −η − (1 − α)δ ρ(ζ − η) − ζ . (33)
1−ρ

96
Online Appendix Allcott, Gentzkow, and Song

E.4 Empirical Moments and Estimation Details

Appendix Table A7 presents the full set of moments and fixed parameter values that are inputs to our
unrestricted model and alternative specifications. In light of the discussion in Section 5.3, we omit the
first half of period 2 when we estimate the anticipatory bonus effect τ2B .22 The average of predicted use with
and without the bonus x̃¯3B,BC and the predicted contemporaneous bonus effect τ̃3B are the predictions before
the bonus MPL on survey 2, as displayed in Figure 7. Because we do not have an explicit elicitation of the
predicted limit effect, we use the actual limit effect τ3L to proxy for the predicted limit effect τ̃3L .23 Since
Figure 6 shows that the average prediction error for period t consumption is similar when elicited on survey
t versus survey t − 1, we let observed Control group misprediction mC proxy for steady-state misprediction
mss .
We winsorize the anticipatory bonus effect at τ2B ≤ 0, which affects 15 percent of draws. We also drop
the 0.32 percent of bootstrap draws in which the denominator of steady-state consumption in equation (15)
is not positive.
22 Appendix Table A8 presents parameter estimates when we use all of period 2 to estimate τ2B . The estimated ρ is larger, as
expected, but the other parameter estimates are very similar.
23 The average difference in predicted FITSBY use between Limit and Limit Control on survey 3 is τ̃ L ≈ −10.5 minutes per day,
3
much smaller than the actual limit effect of τ3L ≈ −22.3 minutes per day. In the limit effect strategy in equation (28), τ̃3L makes
little difference because it is multiplied by (1 − α), which is small. However, in the limit valuation strategy in equation (32), γ̃ is
inversely proportional to τ̃3L , so a much smaller τ̃3L would make the estimated γ̃ much larger.

97
Online Appendix Allcott, Gentzkow, and Song

Table A7: Empirical Moments and Additional Parameters

(1) (2)
Point Confidence
Parameter Description estimate interval
δ Three-week discount factor (unitless) 0.997
τ2B Anticipatory bonus effect (minutes/day) −1.96 [−7.40, 0]
τ3B Contemporaneous bonus effect (minutes/day) -55.9 [-61.7, -50.3]
τ4B Long-term bonus effect (minutes/day) -19.2 [-24.7, -13.7]
τ5B Long-term bonus effect (minutes/day) -12.3 [-18.1, -6.54]
τ2L Limit effect (minutes/day) -24.3 [-28.1, -20.4]
mC , mss Control group misprediction (minutes/day) 6.13 [4.52, 7.72]
x̃¯3B+BC Predicted use with/without bonus (minutes/day) 122 [114, 130]
τ̃3B Predicted bonus effect (minutes/day) −45.0 [−50.0, −40.1]
τ̃3L Predicted limit effect (minutes/day) −22.3 [−27.3, −17.3]
ω Temptation reduction from limit 1
v̄B Average bonus valuation ($/day) 3.20 [3.12, 3.29]
v̄L Average limit valuation ($/day) 0.210 [0.184, 0.237]
pB Bonus price ($/hour) 2.5
F̄ B Average bonus fixed payment ($/day) 7.03 [6.96, 7.09]
x̄1 Average baseline use (minutes/day) 153 [149, 157]
Notes: This table presents point estimates and bootstrapped 95 percent confidence intervals for the empirical moments
used for estimation. We winsorize at τ2B ≤ 0, and we drop the 0.32 percent of draws in which the denominator of
steady-state consumption in equation (15) is not positive.

98
Online Appendix Allcott, Gentzkow, and Song

Table A8: Primary Parameter Estimates Using τ2B for All of Period 2

(1)
Unrestricted
model
Parameter Description (units) (α = α̂)
λ Habit stock effect on consumption (unitless) 1.08
[0.565, 3.09]
ρ Habit formation (unitless) 0.308
[0.113, 0.507]

α Projection bias (unitless) 0.725

[0.427, 0.969]
η Price coefficient ($-day/hour2 ) −2.85
[−3.15, −2.61]
ζ Habit stock effect on marginal utility ($-day/hour2 ) 2.91
[1.49, 8.45]

γ − γ̃ Naivete about temptation ($/hour) 0.283

[0.208, 0.359]
γ Temptation ($/hour) 1.16
[0.938, 1.40]

κ̄ Average intercept ($/hour) −1.95

[−3.40, −0.574]
Notes: This table presents point estimates and bootstrapped 95 percent confidence intervals from the estimation strat-
egy described in Section E.3. We winsorize at τ2B ≤ 0, and we drop the 0.32 percent of draws in which the denominator
of steady-state consumption in equation (15) is not positive. Temptation γ is from the limit effect strategy, using equa-
tion (28). This parallels column 2 of Table 4, except using all of period 2 (instead of only the second half of period 2)
to estimate the anticipatory bonus effect τ2B .

E.5 Alternative Temptation Estimates

Appendix Table A9 presents alternative estimates of temptation γ in the restricted and unrestricted models.
After repeating the primary limit effect estimate, the table reports the bonus valuation estimate. Before the
bonus MPL on survey 2, the average participant predicted that they would use FITSBY 2.5 and 1.6 hours per
day without and with the bonus, respectively. Thus, the average survey taker would have predicted that the
price increase would cause a consumer surplus loss from their period 3 self’s perspective of pB3 x̃¯3 ≈ $2.50 ×
1
2 (2.5 + 1.6) ≈ $5.09 per day of period 3. This is the trapezoid ABCD on Figure 9. The average bonus fixed
payment was F̄ B ≈ $7.03 per day. Thus, if the average participant perceived herself to be time consistent,
she would have been indifferent between the bonus and a certain payment of $7.03 − $5.09 ≈ $1.94 per day.
In reality, the average participant was indifferent between the bonus and a certain payment of $64, or
v̄B ≈ $64/20 ≈ $3.20 per day over the 20-day period. This excess valuation implies a behavior change

99
Online Appendix Allcott, Gentzkow, and Song

premium of $3.20 − $1.94 ≈ $1.26 per day. This is the parallelogram BCEF on Figure 9: the additional
temptation reduction benefit that the period 2 survey taker perceives from the reduced FITSBY use caused
by the bonus. Rearranging this logic into equation (30) gives perceived temptation γ̃ˆ ≈ 1.34 $/hour. Using
− γ̃ ≈ 0.274 gives γ̂ ≈ 1.61 for the bonus valuation strategy in column 1.
the estimated naivete of γ[
The average Limit group participant was indifferent between access to the limit functionality for period
3 and a certain payment of $4.20, or v̄L ≈ $4.20/20 ≈ $0.210 per day over the 20-day period. This is the
triangle on Figure 9: the perceived deadweight loss reduction from the reduced FITSBY use caused by the
L
limit. Inserting this into equation (32) with ω = 1 gives perceived temptation γ̃ˆ = v̄L ≈
−τ̃3 /2
0.210
≈
(−(−22.3)/60)/2
− γ̃ ≈ 0.274 gives γ̂ ≈ 1.41 for the limit valuation strategy in column 1.
1.13 $/hour. Using γ[
So far, we have modeled FITSBY screen time on other devices as part of an outside option that is not
affected by self-control problems. In Appendix G.5, we generalize the model to include multiple temptation
goods. As discussed in Section 5.5, self-reports suggest that the limit increased FITSBY use on other devices
by 4.2 minutes per day, while the bonus reduced FITSBY use on other devices by 8.1 minutes per day. We
use these additional moments to identify the multiple-good model.
The next three rows in Appendix Table A9 present estimates from the multiple-good model. The limit
effect estimate increases to γ̂ ≈ 1.31 $/hour, because in the multiple-good model, more temptation is needed
to explain the observed limits when consumers setting the limits think they’ll evade the limits through
substitution to other devices. The bonus valuation estimate decreases to γ̂ ≈ 1.44 $/hour, because in the
multiple-good model, less temptation is needed to explain the observed bonus valuation when consumers
think the bonus will also reduce FITSBY use on other devices. The limit valuation estimate increases to
to γ̂ ≈ 2.09 $/hour, because in the multiple-good model, more temptation is needed to explain the observed
limit valuation when consumers think the limit will also increase FITSBY use on other devices.
Next, we return to the single-good model and consider an alternative specification where we estimate ω
from differences in self-reported ideal use change between the Limit and Limit Control groups. Intuitively,
if the Limit group reports on survey 3 that looking back over period 2, they ideally would not have further
reduced their screen time, this suggests that the limit functionality fully eliminated temptation (ω = 1).
Extending this intuition, we estimate ω as the share of the Limit Control group’s ideal use change that is
eliminated in the Limit treatment group. If d2g is group g’s average ideal use change reported on survey 3
retrospectively about period 2, this is:
d2L − d2LC
ω= . (34)
−d2LC
In the data, the Limit and Limit Control groups report that they ideally would have changed use by −9.5
−0.095−(−0.15)
and −15 percent, respectively. This gives ω̂ ≈ −(−0.15) ≈ 0.385.
If we assume that the limit only eliminates share ω < 1 of temptation, the limit effect strategy will
deliver larger γ, because we infer that the true effect of temptation on consumption is larger. By contrast, the
limit valuation strategy will deliver smaller γ, because a smaller γ is needed to explain a given valuation v̄L
when temptation has a larger effect on consumption. Appendix Table A9 shows that in the restricted model

100
Online Appendix Allcott, Gentzkow, and Song

(α = 1), the limit effect γ̂ increases from 1.09 to 2.82, while the limit valuation strategy γ̂ decreases from
1.41 to 0.975.
Finally, we extend the limit effect strategy to allow for individual-specific heterogeneity in γ. To do this,
we exploit the facts that we observe each participant’s period 2 limit tightness Hi2 and that tightness is closely
related to the limit treatment effect. We estimate heterogeneous period 2 and 3 limit effects as a function of
period 2 limit tightness by adding an interaction term τ HL Hi2 Li to the treatment effect estimation in equation
(4); see Appendix Table A10.24 For each participant, we insert the fitted limit effect τ̂itL = τ̂tL + τ̂ HL Hi2 into
equation (28) to infer γi . The final row of Appendix Table A9 shows that although this allows substantial
heterogeneity, the average temptation γ̄ is essentially the same as the homogeneous γ from the limit effect
strategy, as one would expect.
These alternative approaches imply temptation γ is between about $1 and $3 per hour. Our primary
strategy (the limit effect) is relatively conservative.
24 H is missing for the Limit Control group, so we are not able to include the main effect of Hi2 in this regression. In theory, this
i
could generate omitted variable bias if period 2 or 3 control group consumption varies with the tightness that they would have set.
Appendix Table A10 shows that Hi2 is associated with the Limit group’s consumption in the second half of period 1 (before the
limit functionality was turned on). However, the association is small compared to the association in periods 2 and 3, which suggests
that the potential omitted variables bias is relatively small.

101
Online Appendix Allcott, Gentzkow, and Song

Table A9: Alternative Temptation Parameter Estimates

(1) (2)
Restricted Unrestricted
model model
Parameter Description (units) B
(τ2 = 0, α = 1) (α = α̂)
γ Temptation ($/hour)
Limit effect (primary) 1.09 1.11
[0.884, 1.30] [0.903, 1.33]
Bonus valuation 1.61 1.62
[1.29, 1.94] [1.29, 1.94]
Limit valuation 1.41 1.41
[1.19, 1.75] [1.19, 1.76]
Limit effect, multiple-good model 1.31
[1.01, 1.71]
Bonus valuation, multiple-good model 1.44 1.45
[1.16, 1.73] [1.17, 1.74]
Limit valuation, multiple-good model 2.09 2.09
[1.33, 7.10] [1.33, 7.10]
Limit effect, ω = ω̂ 2.82 2.92
[2.11, 3.92] [2.22, 4.16]
Limit valuation, ω = ω̂ 0.975 0.979
[0.826, 1.19] [0.833, 1.20]

γ̄ Average temptation ($/hour) 1.08 1.10

Heterogeneous limit effect [0.873, 1.29] [0.889, 1.31]
Notes: This table presents point estimates and bootstrapped 95 percent confidence intervals for alternative estimates
of temptation γ. Each row reflects estimates from a different specification. γ for the limit effect, bonus valuation,
and limit valuation strategies is from equations (28), (30), and (32), respectively, combined with naivete γ − γ̃ from
equation (27). γ for the multiple-good model is from equations (182), (187), and (190) in Appendix G.5; we do not
have a limit effect estimate for the unrestricted multiple-good model. ω̂ is from equation (34).

102
Online Appendix Allcott, Gentzkow, and Song

Table A10: Heterogeneity in Limit Effect by Limit Tightness

2nd half of period 1 Period 2 Period 3

FITSBY use FITSBY use FITSBY use
(1) (2) (3)
Bonus treatment −4.702 −3.228 −54.384
(2.001) (2.154) (2.835)

Limit treatment −5.281 0.447 −1.248

(2.143) (2.308) (3.041)

Limit treatment × period 2 limit tightness 0.114 −0.551 −0.469

(0.027) (0.029) (0.039)

1st half of period 1 FITSBY use 0.845

(0.014)

Period 1 FITSBY use 0.894 0.795

(0.015) (0.020)

Observations 1,933 1,930 1,931

R2 0.849 0.795 0.665

Notes: This table presents the effects of bonus and limit treatments on FITSBY use in periods 1, 2, and 3 using
equation (4), including an additional interaction between the Limit group indicator and period 2 limit tightness. Limit
tightness is the amount by which a user’s limits would have hypothetically reduced overall screen time if applied
to their baseline use without snoozes; see equation (5). FITSBY use refers to screen time on Facebook, Instagram,
Twitter, Snapchat, browser, and YouTube.

103
Online Appendix Allcott, Gentzkow, and Song

E.6 Model Estimates with Sample Weights

Table A11: Demographics in Weighted Sample

(1) (2) (3)

Analysis Balanced U.S.
sample sample adults
Income ($000s) 40.8 42.1 43.0
College 0.67 0.55 0.30
Male 0.39 0.42 0.49
White 0.72 0.72 0.74
Age 33.7 38.7 47.6
Period 1 phone use (minutes/day) 333.0 339.3 .
Period 1 FITSBY use (minutes/day) 152.8 155.4 .

Notes: Column 1 presents average demographics for our analysis sample, column 2 presents average demographics
for our weighted sample, and column 3 presents average demographics of American adults using data from the 2018
American Community Survey. The sample weights are initially calculated to make the sample nationally representative
on these five demographics but are then winsorized at [1/3, 3] to reduce precision loss.

104
Online Appendix Allcott, Gentzkow, and Song

Table A12: Empirical Moments and Additional Parameters in Weighted Sample

(1) (2)
Point Confidence
Parameter Description estimate interval
δ Three-week discount factor (unitless) 0.997
τ2B Anticipatory bonus effect (minutes/day) −4.41 [−12.8, 0]
τ3B Contemporaneous bonus effect (minutes/day) -58.5 [-67.3, -50.3]
τ4B Long-term bonus effect (minutes/day) -25.1 [-34.7, -15.7]
τ5B Long-term bonus effect (minutes/day) -16.4 [-26.5, -7.93]
τ2L Limit effect (minutes/day) -23.3 [-29.6, -16.7]
mC Control group misprediction (minutes/day) 4.96 [3.03, 7.11]
x̃¯3B+BC Predicted use with/without bonus (minutes/day) 127 [114, 140]
τ̃3B Predicted bonus effect (minutes/day) −49.4 [−56.6, −41.9]
τ̃3L Predicted limit effect (minutes/day) −20.7 [−27.9, −12.7]
ω Temptation reduction from limit 1
v̄B Average bonus valuation ($/day) 3.29 [3.15, 3.44]
v̄L Average limit valuation ($/day) 0.271 [0.229, 0.315]
pB Bonus price ($/hour) 2.5
F̄ B Average bonus fixed payment ($/day) 6.84 [6.72, 6.96]
x̄1 Average baseline use (minutes/day) 156 [149, 164]
Notes: This table presents point estimates and bootstrapped 95 percent confidence intervals for the empirical moments
used for estimation. We winsorize at τ2B ≤ 0, and we drop the 0.32 percent of draws in which the denominator of
steady-state consumption in equation (15) is not positive. This parallels Table 3, except using the weighted sample.
The sample weights are initially calculated to make the sample nationally representative on the five demographics in
Appendix Table A11 but are then winsorized at [1/3, 3] to reduce precision loss.

105
Online Appendix Allcott, Gentzkow, and Song

Table A13: Model Parameter Estimates in Weighted Sample

Restricted
model
Parameter Description (units) (τ2B = 0, α = 1)
λ Habit stock effect on consumption (unitless) 1.93
[0.757, 3.89]
ρ Habit formation (unitless) 0.223
[0.122, 0.469]

α Projection bias (unitless) 1

η Price coefficient ($-day/hour2 ) −2.57

[−2.98, −2.23]
ζ Habit stock effect on marginal utility ($-day/hour2 ) 4.95
[2.07, 9.96]

γ − γ̃ Naivete about temptation ($/hour) 0.212

[0.130, 0.307]
γ Temptation ($/hour) 0.998
[0.709, 1.30]

κ̄ Average intercept ($/hour) −1.99

[−3.52, −0.422]
Notes: This table presents point estimates and bootstrapped 95 percent confidence intervals from the estimation strat-
egy described in Section E.3. We winsorize at τ2B ≤ 0, and we drop the 0.32 percent of draws in which the denominator
of steady-state consumption in equation (15) is not positive. Temptation γ is from the limit effect strategy, using equa-
tion (28). This parallels Table 4, except using the weighted sample. The sample weights are initially calculated to
make the sample nationally representative on the five demographics in Appendix Table A11 but are then winsorized at
[1/3, 3] to reduce precision loss.

F Proofs of Propositions in Appendix E.1

Given naivete about projection bias, the predicted continuation value function given predicted consumption
and habit stock is

T
Vt+1 (s̃t+1 ) = ∑ δ r−t ur (x̃r∗ (s̃r , γ̃, p r ) ; s̃r , pr ) . (35)
r=t+1

The consumer’s predicted objective function in future period t can thus be written as

Ũt (xt ; s̃t ) = ut (xt ; s̃t , pt ) + γ̃xt + δVt+1 (s̃t+1 ) , (36)

106
Online Appendix Allcott, Gentzkow, and Song

and the consumer’s actual period t objective function from equation (2) can be written as

α ∑Tr=t+1 δ r−t ur (x̃r∗ (st , γ̃, p r ) ; st , pr )

Ut (xt ; st ) = ut (xt ; st , pt ) + γxt + . (37)
+(1 − α)δVt+1 (s̃t+1 )

Recall that we defined ut := ut (xt∗ ; st , pt ), x̃r := x̃r∗ (s̃r , γ̃, p r ), and ũr := ur (x̃r ; s̃r , pr ).

F.1 Proof of Proposition 1: Euler Equation

In this section, we derive the Euler equation (equation (18)), proving Proposition 1.

Proof. The time t first-order condition from maximizing utility (equation (37)) is

∂ ut d s̃t+1 dVt+1 (s̃t+1 )

+ γ = −(1 − α)δ (38)
∂ xt dxt d s̃t+1
 
d s̃t+1 ∂ ũt+1 ∂ x̃t+1 ∂ ũt+1 d s̃t+2 dVt+2 (s̃t+2 )
= −(1 − α)δ + − (1 − α)δ 2 (39)
dxt+1 ∂ x̃t+1 ∂ s̃t+1 ∂ s̃t+1 dxt d s̃t+2
   
∂ ũt+1 ∂ x̃t+1 ∂ ũt+1 2 ∂ x̃t+1 dVt+2 (s̃t+2 )
= −(1 − α)δ ρ + − (1 − α) (δ ρ) 1 + , (40)
∂ x̃t+1 ∂ s̃t+1 ∂ s̃t+1 ∂ s̃t+1 d s̃t+2

where the third line uses the fact that the total derivative of predicted period t + 2 habit stock with respect to
period t consumption is

d s̃t+2 ∂ s̃t+2 ∂ s̃t+1 ∂ s̃t+2 ∂ x̃t+1 ∂ s̃t+1

= +
dxt ∂ s̃t+1 ∂ xt ∂ x̃ ∂ s̃ ∂ xt
  t+1 t+1
∂ x̃t+1
= ρ2 1 + (41)
∂ s̃t+1

The time t self predicts that the time t + 1 self will maximize equation (36), setting xt+1 according to
the following first-order condition:

∂ ũt+1 d s̃t+2 dVt+2 (s̃t+2 )

0= + γ̃ + δ (42)
∂ x̃t+1 dxt+1 d s̃t+2
∂ ũt+1 dVt+2 (s̃t+2 )
= + γ̃ + δ ρ (43)
∂ x̃t+1 d s̃t+2
 
Multiplying the predicted time t + 1 first-order condition by (1 − α)δ ρ 1 + ∂∂ x̃s̃t+1
t+1
gives

    
∂ x̃t+1 ∂ ũt+1 ∂ x̃t+1 dVt+2 (s̃t+2 )
0 = (1 − α)δ ρ 1 + + γ̃ + (1 − α) (δ ρ)2 1 + (44)
∂ s̃t+1 ∂ x̃t+1 ∂ s̃t+1 d s̃t+2

107
Online Appendix Allcott, Gentzkow, and Song

The last term is the same as the last term in the time t first-order condition. Adding this equation to the
time t first-order condition yields

    
∂ ut ∂ x̃t+1 ∂ ũt+1 ∂ ũt+1 ∂ x̃t+1 ∂ ũt+1
+ γ = (1 − α)δ ρ 1 + + γ̃ − (1 − α)δ ρ + (45)
∂ xt ∂ s̃t+1 ∂ x̃t+1 ∂ x̃t+1 ∂ s̃t+1 ∂ s̃t+1
 
∂ ut ∂ ũt+1 ∂ x̃t+1 ∂ ũt+1
+ γ = (1 − α)δ ρ + γ̃ + γ̃ − . (46)
∂ xt ∂ x̃t+1 ∂ s̃t+1 ∂ s̃t+1

We now derive the Euler equation with our quadratic functional form. The partial derivatives are

∂ ut
=ηxt∗ + ζ st + ξt − pt (47)
∂ xt
∂ ũt+1
=η x̃t+1 + ζ s̃t+1 + ξt+1 − pt+1 (48)
∂ x̃t+1
∂ x̃t+1
λ̃t+1 := (49)
∂ s̃t+1
∂ ũt+1
=ζ x̃t+1 + φ . (50)
∂ s̃t+1

Substituting these into equation (46) yields equation (18).

F.2 Proof of Proposition 2: Linear Policy Functions

In this section, we first show that the policy function is linear in habit stock. We then show that if the
objective function is concave, λ converges to a constant far from the time horizon. We then show the
conditions under which utility is concave. Finally, we show the condition required for µ to converge to a
constant far from the time horizon. Our proof strategy follows Gruber and Köszegi (2001).

Lemma 2. Suppose ut (xt ; st , pt ) is given by equation (3) and (x0∗ , ..., xT∗ ) is a perception-perfect strategy
profile. Then for any t,

xt∗ (st , γ, pt ) = λt st + µt (γ) (51)

x̃t∗ (st , γ̃, pt ) = λ̃t st + µt (γ̃) (52)

where λt is a function of only {η, ζ , δ , ρ, α}, λ̃t is a function of only {η, ζ , δ , ρ}, and µt is linear in pt .

Proof. We prove by backwards induction. First, we show that the result holds for period T . Given our
functional form, the period T first-order condition is

108
Online Appendix Allcott, Gentzkow, and Song

ηxT∗ + ζ sT + ξT − pT + γ = 0, (53)

and thus
ζ sT + ξT − pT + γ
xT∗ = . (54)
−η
Thus, xT∗ can be written as

xT∗ = λT sT + µT (γ), (55)

ζ ξT −pT +γ
with λT = −η and µT (γ) = −η .
Analogously, predicted consumption is

ζ s̃T + ξT − pT + γ̃
x̃T = , (56)
−η
so x̃T can be written as

x̃T = λT s̃T + µT (γ̃), (57)

ξT −pT +γ̃
with µT (γ̃) = −η . The function µT is linear in pT .
Now, we use the Euler equation to show that if the result holds for t + 1, it holds for t. The Euler
equation is

   
∂ x̃t+1
ηxt∗ + ζ st + ξt − pt + γ = (1 − α)δ ρ η x̃t+1 + ζ s̃t+1 + ξt+1 − pt+1 + 1 + γ̃ − ζ x̃t+1 − φ
∂ s̃t+1
   
∂ x̃t+1
= (1 − α)δ ρ (η − ζ )x̃t+1 + ζ s̃t+1 + ξt+1 − pt+1 + 1 + γ̃ − φ
∂ s̃t+1
∗
∂ x̃t+1
Substituting x̃t+1 = λ̃t+1 s̃t+1 + µt+1 (γ̃), s̃t+1 = ρ (st + xt∗ ), and λ̃t+1 = ∂ s̃t+1 gives

h   i
ηxt∗ +ζ s̃t +ξt − pt +γ = (1−α)δ ρ (η − ζ ) λ̃t+1 ρ (xt∗ + st ) + µt+1 (γ̃) + ζ ρ(st + xt∗ ) + ξt+1 − pt+1 + γ̃ + γ̃ λ̃t+1 − φ .
(58)
Solving for xt∗ gives

h  i h i
st ζ − (1 − α)δ ρ 2 (η − ζ )λ̃t+1 + ζ + ξt − pt + γ − (1 − α)δ ρ (η − ζ )µt+1 (γ̃) + ξt+1 − pt+1 + γ̃ + γ̃ λ̃t+1 − φ
xt∗ =   .
−η + (1 − α)δ ρ 2 (η − ζ )λ̃t+1 + ζ
(59)

109
Online Appendix Allcott, Gentzkow, and Song

Thus, xt∗ = λt st + µt (γ), with

 
ζ − (1 − α)δ ρ 2 (η − ζ )λ̃t+1 + ζ
λt =  , (60)
−η + (1 − α)δ ρ 2 (η − ζ )λ̃t+1 + ζ

and
h i
ξt − pt + γ − (1 − α)δ ρ ξt+1 − pt+1 + γ̃ + γ̃ λ̃t+1 − φ + (1 − α)δ ρ (ζ − η) µt+1 (γ̃)
µt (γ) =   . (61)
−η + (1 − α)δ ρ 2 (η − ζ )λ̃t+1 + ζ

We can analogously begin with the period t Euler equation as predicted before period t, which has γ̃ and
s̃t instead of γ and st on the left-hand side, and does not have the (1 − α) term. This gives x̃t = λ̃t s̃t + µt (γ̃),
with  
ζ − δ ρ 2 (η − ζ )λ̃t+1 + ζ
λ̃t =  . (62)
−η + δ ρ 2 (η − ζ )λ̃t+1 + ζ

and µt (γ̃) given by equation (61) except that, as as implied by writing µt (γ̃) instead of µt (γ), the third term
in the numerator is γ̃ instead of γ.25 Thus, λt is not correctly perceived in advance of period t.
λt depends only on {η, ζ , δ , ρ, α}, and λ̃t depends only on {η, ζ , δ , ρ}, as long as λ̃t+1 depends only
on {η, ζ , δ , ρ}. Because consumers misperceive γ, µr is also misperceived for r > t. The function µt is
linear in pt .

We now show that with concave utility, λt and λ̃t are constant in t far from the time horizon.

Lemma 3. Suppose the conditions for Lemma 2 hold and utility is concave. Then for any fixed t,
 
ζ − (1 − α)δ ρ 2
(η − ζ )λ̃ + ζ
λ = lim λt =  , (63)
T →∞
−η + (1 − α)δ ρ 2 (η − ζ )λ̃ + ζ

and
q
ρ 2 (ζ −η)
−η − η 2 − 4 δ(1−δ ρ 2)
ζ
λ̃ = lim λ̃t = 2δ ρ 2 (ζ −η)
. (64)
T →∞
(1−δ ρ 2 )

Proof. To show that λt is constant in t far from the time horizon, it suffices to prove the convergence of λ̃t
to the steady state, since λt is a function of λ̃t+1 and other deterministic parameters. We define the function
25 Equation (60) is much simpler than equation (25) of Gruber and Köszegi (2001), and our expression for λ does not depend on
t
actual or perceived temptation γ or γ̃, while theirs depends on present focus β . This is because in their quasi-hyperbolic framework,
1 − β multiplies λt+1 parameters in the Euler equation and doesn’t drop out.

110
Online Appendix Allcott, Gentzkow, and Song

f (λ̃ ) according to Equation (62) that describes the recursion λ̃t = f (λ̃t+1 ). We first find the values of λ̃ that
could be fixed points. Assuming constant λ̃ and rearranging Equation (60) gives
 
−η λ̃ + δ ρ 2 (η − ζ ) λ̃ 2 + ζ λ̃ = ζ + δ ρ 2 ((ζ − η) − ζ ) . (65)

Collecting terms gives

λ̃ 2 δ ρ 2 (η − ζ ) + λ̃ η δ ρ 2 − 1 + ζ δ ρ 2 − 1 = 0



(66)
δ ρ 2 (ζ
− η)
λ̃ 2 + λ̃ η + ζ = 0. (67)
(1 − δ ρ 2 )

Using the quadratic formula gives

q
ρ 2 (ζ −η)
−η ± η 2 − 4 δ(1−δ ρ 2)
ζ
λ̃ = 2δ ρ 2 (ζ −η)
. (68)
(1−δ ρ 2 )

We now prove convergence. The function f (λ ) has the following properties. First, f (λ ) is always
increasing as

     
−δ ρ 2 (η − ζ ) −η + δ ρ 2 (η − ζ ) λ̃ + ζ − δ ρ 2 (η − ζ ) ζ − δ ρ 2 (η − ζ ) λ̃ + ζ
f 0 (λ̃ ) =   2 (69)
−η + δ ρ 2 (η − ζ ) λ̃ + ζ
δ ρ 2 (ζ − η)2
=  2 > 0. (70)
2
−η + δ ρ (η − ζ ) λ̃ + ζ

−η + δ ρ 2 ζ
Second, f is convex on (−∞, λ̃¯ ),where λ̃¯ = > 0. This comes from the sign of its second
δ ρ 2 (ζ − η)
derivative
2δ 2 ρ 4 (−η + ζ )3
f 00 (λ̃ ) =   3 , (71)
−η + δ ρ 2 (η − ζ ) λ̃ + ζ

which is determined by the sign of the denominator.

Third, for λ̃ > λ̃¯ , f (λ̃ ) is always negative due to the denominator in equation (62), hence none of the
solutions for a constant λ̃t are in this region.
Fourth, f (0) > 0 since δ ρ 2 < 1 and

ζ (1 − δ ρ 2 )
f (0) = . (72)
−η + δ ρ 2 ζ

111
Online Appendix Allcott, Gentzkow, and Song

Fifth, f (λ̃ ) is continuous on [0, λ̃¯ ) and limλ̃ →λ̃¯ f (λ̃ ) = ∞ as the denominator in equation (60) goes to
0.
The properties highlighted above imply that both candidate solutions for a constant λ̃t in equation (68)
are positive. To see this, denote the two candidate solutions as (λ̃1 , λ̃2 ), with λ̃1 < λ̃2 . Since f (0) > 0, we
know that at least one solution for λ̃ is positive given −η > 0. Furthermore, since f (λ̃ ) > 0 on (−∞, λ̃¯ ], it
cannot be true that an increasing, continuous, and convex function that diverges to infinity at λ̃¯ only crosses
the identity function once in [0, λ̃¯ ). Hence, both solutions are in [0, λ̃¯ ].
Given this result and the convex shape of this function, it must be true that λ̃1 is a stable constant
solution for the recursion while λ̃2 is unstable. For any point in [0, λ̃1 ] the recursion implies an increase in
λ̃t ( f (λ̃ ) > λ̃ ), for any point in [λ̃1 , λ̃2 ] the recursion implies a decrease in λt ( f (λ̃ ) < λ̃ ), and for any point
in [λ̃ , λ̃¯ ] the recursion implies an increase in λ̃ ( f (λ̃ ) > λ̃ ). Overall, this means that for any starting value
2 t
of λ̃t ∈ [0, λ̃2 ) the recursion converges to λ̃1 .
To complete the proof, we begin with λ̃T and then prove that far away from the time horizon, λ̃t is
ζ
constant. To do this, we need to show that this initial value, given by λ̃T = , is less than λ̃2 . To show
−η
−η(1 − δ ρ 2 )
this, notice that the two solutions (λ̃1 , λ̃2 ) are symmetrically placed around λ˜s = . Given
2δ ρ 2 (ζ − η)
this value, by the parametric assumption that guarantees the existence of the two constant solutions for the
recursion, we know that
δ ρ 2 (ζ − η)
η2 − 4 ζ > 0, (73)
(1 − δ ρ 2 )
and since
δ ρ 2 (ζ − η) ζ ζ −η(1 − δ ρ 2 )
η2 > 2 ⇐⇒ < , (74)
(1 − δ ρ̃ 2 ) −η 2δ ρ 2 (ζ − η)
we have that λ̃T < λ˜s . Then λ̃T < λ˜s < λ2 , and hence the backward recursion starting from λ̃T converges
ζ −X
far from the time horizon to a stationary value λ̃ ∗ = λ̃1 . Moreover, f (λ̃T ) can be written as −η+X , and we
ζ −X ζ
know that −η+X > −η whenever X < 0. Then, given that

  ζ ζ2
X = (1 − α)δ ρ 2 (η − ζ ) λ̃T + ζ < 0 ⇐⇒ (η − ζ ) + ζ < 0 ⇐⇒ < 0 ⇐⇒ η < 0,
−η η

we have X < 0. Thus we can conclude that f (λ̃T ) > λ̃T and therefore, λ̃T < λ̃1 . Thus, we have proved
that the backward recursion converges to an stationary value of λ̃ ∗ = λ̃1 , and it does so as an increasing
sequence.
Finally, we demonstrate that λt also converges to a steady-state in a decreasing manner. We note that
 
ζ − (1 − α)δ ρ 2 (η − ζ )λ̃ + ζ
λ = g(λ̃ ) =   (75)
−η + (1 − α)δ ρ 2 (η − ζ )λ̃ + ζ

112
Online Appendix Allcott, Gentzkow, and Song

Which we can rewrite as

 
ζ + (1 − α)δ ρ 2 (ζ − η)λ̃ + ζ
λ = g(λ̃ ) =   (76)
−η − (1 − α)δ ρ 2 (ζ − η)λ̃ + ζ

Note that (1 − α)δ ρ 2 (ζ − η) is positive, so the numerator decreases when λ̃ decreases, whereas the
denominator increases, since −(1−α)δ ρ 2 (ζ −η)λ̃ becomes less negative. Hence, g(λ̃ ) = λ also decreases
when λ̃ decreases.

We now show that utility is concave in xt as long as there is not too much habit formation in a specific
sense.

dUt
Lemma 4. Suppose the conditions for Lemma 2 hold and Ut is given by equation (37). Then for any t, dxt
(1−α)λ̃ b −η ∂ 2Ut
is continuous in xt . Furthermore, if λ̃ b is an upper bound on λ̃t and 2 < ζ
, then ∂ xt2
<0
( 1+λ̃ b )−δ ρ 2 (
1+λ̃ b )
for all t ≥ 0.

Proof. The period t decisionmaker maximizes equation (37). The derivative of equation (37) can be written
as

 
T
dUt (xt ; st ) ∂ ut r−t ∂ s̃r 
 ∂ ũr ∂ x̃r ∂ ũr ∂Vr+1 ∂ x̃r 
= + γ + (1 − α) ∑ δ + + . (77)

 δρ
dxt ∂ xt ∂ xt  ∂ x̃ ∂ s̃r ∂ s̃r ∂ s̃r+1 ∂ s̃r
r=t+1 | r {z

} | {z }
effect of s̃r on period r utility partial effect on future utility

The summation term in equation (77) is the effect on future utility from the change in habit stock brought
∂ s̃r
into future periods. ∂ xt = ρ r−t is the predicted direct effect of consumption x̃t on stock in period r. The
first two terms inside brackets are the effect of that change on period r utility. The final term inside brackets
accounts for the fact that the resulting change in x̃r will affect utility in later periods.
The period t decisionmaker predicts that her period r > t self will maximize equation (36). The pre-
dicted period r first-order condition is

dŨr (xr ; s̃r )  ∂ ũr ∂Vr+1
d x̃r  = 0 = ∂ x̃r + γ̃ + δ ρ ∂ s̃r+1 . (78)
x̃r

∂ x̃r
Multiplying this FOC by λ̃r := ∂ s̃r and subtracting it from the term inside brackets in equation (77)
gives

113
Online Appendix Allcott, Gentzkow, and Song

 
∂ ũr
dUt ∂ ut T λ̃ + ∂ ũr + δ ρ ∂V r+1
∂ s̃r+1 λ̃r i
= + γ + (1 − α) ∑ δ r−t ρ r−t  h∂ x̃r r ∂ s̃r  (79)
dxt ∂ xt r=t+1 − ∂∂ ũx̃rr λ̃r + γ̃ λ̃r + δ ρ ∂V r+1
∂ s̃r+1 λ̃r
T  
∂ ut r−t ∂ ũr
= + γ + (1 − α) ∑ (δ ρ) − γ̃ λ̃r (80)
∂ xt r=t+1 ∂ s̃r

With the quadratic functional form, this becomes

T
dUt h i
= ηxt + ζ st + ξt − pt + γ + (1 − α) ∑ (δ ρ) r−t ζ x̃r + φ − γ̃ λ̃ . (81)
dxt r=t+1

In this equation, two terms (xt and x̃r ) depend on xt . xt is by definition continuous in xt , and x̃r is
dUt
continuous in past consumption xt due to the evolution of habit stock and Lemma 2. Thus, dxt is continuous
in x.
We now turn to concavity. The derivative of equation (81) is

d 2Ut ∞
d x̃r
2
= η + (1 − α) ∑ (δ ρ) r−t ζ . (82)
dxt r=t+1 dxt
" #
∞ r−1  
r−t r−t
= η + (1 − α) ∑ (δ ρ) ζ λ̃r ρ ∏ 1 + λ̃ j (83)
r=t+1 j=t+1

d 2Ut
Intuitively, dxt2
< 0 requires that the diminishing marginal utility in period t outweighs the incentive to
increase current consumption for the purpose of increasing future utility through ζ . This will tend to be true
when projection bias α is large and/or habit formation ρ is small. A small ρ has a direct effect by causing
d x̃r
the habit stock from dxt to decay faster. It also has an indirect effect by reducing dxt , the perceived effect of
current consumption on future consumption.
If we know an upper bound λ̃ b such that λ̃ b > λ̃t for all t, we can write a simpler necessary condition

114
Online Appendix Allcott, Gentzkow, and Song

d 2Ut
for concavity: dxt2
< 0 for all t ≥ 0 if
" #
∞ r−1   −η
(1 − α) ∑ (δ ρ) r−t λ̃r ρ r−t ∏ 1 + λ̃ j < (84)
r=t+1 j=t+1 ζ
∞   r−t−1  −η
(1 − α) ∑ (δ ρ) r−t λ̃ b ρ r−t 1 + λ̃ b < (85)
r=t+1 ζ
λ̃ b ∞    −η
(1 − α) · ∑ δ ρ 2 1 + λ̃ b r−1 < (86)
b
1 + λ̃ r=1 ζ
 
λ̃ b  1 −η
(1 − α) ·     < (87)
1 + λ̃ b ζ
1 − δ ρ 2 1 + λ̃ b
(1 − α)λ̃ b −η
   2 < . (88)
ζ
1 + λ̃ b − δ ρ 2 1 + λ̃ b

From the proof of Lemma 3, we know that λ̃t decreases as t → T .

Finally, we show the conditions under which µt converges to a constant far from the time horizon.
h   i
Lemma 5. Suppose the conditions for Lemma 2 hold, and −η > (1 − α)δ ρ (ζ − η) 1 + ρ λ̃t+1 − ρζ .
Then lim(T −t)→∞ µt = µ.

Proof. Since µt (γ) is a function of only constants, λ̃t+1 (which converges per Lemma 3), and µt+1 (γ̃), it is
sufficient to show that the sequence µt (γ̃) converges. The coefficient on µt+1 (γ̃) in equation (61) is

(1 − α)δ ρ (ζ − η)
 . (89)
2
−η + (1 − α)δ ρ (η − ζ )λ̃t+1 + ζ

The sequence µt+1 (γ̃) will converge if and only if

(1 − α)δ ρ (ζ − η)
  < 1. (90)
−η + (1 − α)δ ρ 2 (η − ζ )λ̃t+1 + ζ

The denominator is positive at our parameter values, so this inequality requires

h   i
−η > (1 − α)δ ρ (ζ − η) 1 + ρ λ̃t+1 − ρζ . (91)

In words, this requires that perceived habit formation (1 − α)ρ is small relative to the demand slope param-
eter η.

115
Online Appendix Allcott, Gentzkow, and Song

Proposition 2 combines Lemmas 2, 3, 4, and 5.

F.3 Proof of Lemma 1: Steady-State Convergence

Proof. Capital stock evolves according to st = ρ (st−1 + xt−1 ). Substituting in the stable equilibrium strategy
xt∗ = λ st + µ gives

st = ρ (st−1 + λ st−1 + µ) (92)

= ρ µ + ρ (1 + λ ) st−1 (93)
= ρ µ + ρ (1 + λ ) (ρ µ + ρ (1 + λ ) st−2 ) (94)
2 2 2
= ρ µ + ρ (1 + λ ) µ + ρ (1 + λ ) st−2 (95)
2 3 2 3 3
= ρ µ + ρ (1 + λ ) µ + ρ (1 + λ ) µ + ρ (1 + λ ) st−3 . (96)

Thus
µ  
st = ι + ι 2 + ... + ι k + ι k st−k , (97)
1+λ
where ι = (1 + λ ) ρ. Thus, provided that ι < 1, in the limit as k → ∞ we have

µ ι
st = · (98)
1+λ 1−ι
µρ
= . (99)
1 − (1 + λ ) ρ

We can then check that this is indeed a steady state:

  
µρ µρ
st = ρ +µ +λ (100)
1 − (1 + λ ) ρ 1 − (1 + λ ) ρ
 
µρ + µ (1 − (1 + λ ) ρ) + λ µρ
=ρ (101)
1 − (1 + λ ) ρ
 
µρ + µ − µρ − µλ ρ + λ µρ
=ρ (102)
1 − (1 + λ ) ρ
µρ
= (103)
1 − (1 + λ ) ρ

F.4 Proof of Proposition 3: Steady-State Consumption

Proof. We assume steady state implies constant consumption and habit stock, but not necessarily constant
predicted consumption and habit stock. In steady state, pt = p, ξt = ξ , st = sss , and xt = xss . By equation (1)

116
Online Appendix Allcott, Gentzkow, and Song

ρ
governing the evolution of habit stock, sss = ρ(sss + xss ), and re-arranging this equation gives sss = 1−ρ xss .
Earlier, we defined steady-state misprediction as mss := x̃t+1 − xss .
We substitute pt = p, ξt = ξ , st = sss , and xt = xss into the Euler equation (equation (18)), giving

h   i
ηxss + ζ sss + ξ − p + γ = (1 − α)δ ρ η x̃t+1 + ζ ρ (xss + sss ) + ξ − p + 1 + λ̃ γ̃ − ζ x̃t+1 − φ . (104)

ρ
Substituting in sss = 1−ρ xss and also writing predicted consumption as a deviation from the actual value
gives
   
ρζ 1  
ηxss + ξ − p + xss + γ =(1 − α)δ ρ (η − ζ ) ((x̃t+1 − xss ) + xss ) + ζ ρ xss + ξ − p + 1 + λ̃ γ̃ − φ .
1−ρ 1−ρ
(105)
Substituting mss := x̃t+1 − xss and collecting terms gives

   
ρζ ζρ
xss η + − (1 − α)δ ρ (η − ζ ) + =p − ξ − γ + (1 − α)δ ρ (η − ζ )mss
1−ρ 1−ρ
  
+ ξ − p + 1 + λ̃ γ̃ − φ (106)

ρ − (1 − α)δ ρ 2
  
xss η − (1 − α)δ ρ(η − ζ ) + ζ = (1 − (1 − α)δ ρ) (p − ξ ) + (1 − α)δ ρ (η − ζ )mss
1−ρ
  
+ 1 + λ̃ γ̃ − φ − γ. (107)

Multiplying both sides by (−1), setting κ := (1 − α)δ ρ(φ − ξ ) + ξ , and dividing through gives equa-
tion (21).

G Derivations of Estimating Equations in Appendix E.3

We define yg := Ei∈g yi as the expectation over individuals in group g of parameter y. Due to random as-
0 0
signment, ξtg = ξtg and sg2 = sg2 for all {g, g0 }, and µtB = µtBC for t ∈ {2, 4, 5}. The estimating equations for
the restricted model in Section 6.2 are the below equations with the additional assumptions that τ2B = 0 and
α = 1.

G.1 Habit Formation

Derivation of equation (22). From equation (19) and the evolution of habit stock, we have

117
Online Appendix Allcott, Gentzkow, and Song

x4∗ = λ s4 + µ4 (108)
= λ ρ (s3 + x3∗ ) + µ4 (109)
= λ ρ (ρ (s2 + x2∗ ) + x3∗ )+µ4 . (110)

Thus, group average consumption is x4g = λ ρ 2 sg2 + x2g + ρx3g + µ4g , and the period 4 bonus effect is

τ4B = λ ρ 2 τ2B + ρτ3B .

(111)

Re-arranging gives equation (22).

Derivation of equation (23). Similarly, we have

x5∗ = λ s5 + µ5 (112)
= λ ρ (s4 + x4∗ ) + µ5 (113)
= λ ρ (ρ (s3 + x3∗ ) + x4∗ ) + µ5 (114)
= λ ρ (ρ (ρ (s2 + x2∗ ) + x3∗ ) + x4∗ ) + µ5 . (115)

Thus, group average consumption is x5g = λ ρ 3 sg2 + x2g + ρ 2 x3g + ρx4g + µ5g , and the period 5 bonus

effect is
τ5B = λ ρ 3 τ2B + ρ 2 τ3B + ρτ4B .


(116)

Multiplying equation (111) by ρ and subtracting from equation (116) gives τ5B − τ4B ρ = λ ρτ4B , and
re-arranging gives equation (23).

System of equations for λ and ρ. Re-arranging equation (23) gives

τ5B
λ= − 1. (117)
τ4B ρ

Substituting this into equation (22) gives:

τ5B − τ4B ρ τ4B

= (118)
τ4B ρ ρτ3B + ρ 2 τ2B

2
τ4B = τ5B − τ4B ρ τ3B + ρτ2B



(119)
h
i
2
0 = τ2B τ4B ρ 2 + τ3B τ4B − τ2B τ5B ρ + τ4B − τ3B τ5B .
   
(120)

118
Online Appendix Allcott, Gentzkow, and Song

The quadratic formula gives

 r i
B B B B
 B B B B
2  B B  h B
2 B B
− τ3 τ4 − τ2 τ5 ± τ3 τ4 − τ2 τ5 − 4 τ2 τ4 τ4 − τ3 τ5
ρ=   . (121)
2 τ2B τ4B

In all bootstrap draws in our data, only one of the two solutions satisfies the requirement that ρ ≥ 0.
Special case with τ2B = 0. If there is no anticipatory demand response (τ2B = 0), we have τ4B = λ ρτ3B
and τ5B = λ ρ 2 τ3B + λ ρτ4B . Dividing the two equations gives

τ5B τ4B
= ρ +
τ4B τ3B
τ5B τ4B
ρ= − . (122)
τ4B τ3B

We then solve for λ by inserting this ρ into equation (22) with τ2B = 0.

G.2 Perceived Habit Formation, Price Response, and Habit Stock Effect on Marginal Util-
ity
The expectation over i of the Euler equations for group g is

h
i
ηxtg + ζ stg + ξtg − pt + γ = (1 − α)δ ρ η x̃t+1
g g
+ ζ s̃t+1 g
+ ξt+1 g
− pt+1 + γ̃ + γ̃ λ̃t+1 − ζ x̃t+1 +φ . (123)

Derivation of equation (24). Differencing the Euler equations for periods 2 versus 3 for the Bonus
and Bonus Control groups gives

ητ2B = (1 − α)δ ρ −pB + (η − ζ ) x̃3B − x̃3BC + ζ s̃B3 − s̃BC




3 . (124)

Substituting x̃3B − x̃3BC = τ̃3B and s̃B3 − s̃BC B

3 = ρτ2 gives

ητ2B = (1 − α)δ ρ −pB + (η − ζ )τ̃3B + ζ ρτ2B .

 
(125)

Rearranging gives equation (24).

If γ̃ 6= γ, then people update their predictions of x̃3 as they set x2∗ , and thus the predictions of x̃3 from
survey 2 are inconsistent with x2∗ . However, there is only limited misprediction in our data, so this is not
very consequential.
Derivation of equation (25). Differencing the Euler equations for periods 3 versus 4 for the Bonus
and Bonus Control groups gives

119
Online Appendix Allcott, Gentzkow, and Song

−pB − 0 + ητ3B + ζ sB3 − sBC = (1 − α)δ ρ (η − ζ ) x̃4B − x̃4BC + ζ s̃B4 − s̃BC




3 4 . (126)

Habit stock evolution implies sB3 −sBC B BC B BC B B BC B BC B BC =

3 = ρ(s2 −s2 +x2 −x2 ) = ρτ2 and s̃4 − s̃4 = ρ s3 − s3 + x3 − x3
ρ ρτ2B + τ3B . Linear policy functions imply x̃4 = λ̃ s̃4 + µ̃4 , so x̃4B − x̃4BC = λ̃ s̃B4 − s̃BC



4 . Substituting these
equations gives
h 
i
−pB − 0 + ητ3B + ζ ρτ2B = (1 − α)δ ρ (η − ζ )λ̃ + ζ ρ ρτ2B + τ3B .


(127)

Rearranging gives


  
η τ3B − (1 − α)δ ρ 2 λ̃ ρτ2B + τ3B = pB − ζ ρτ2B + (1 − α)δ ρ 2 ζ 1 − λ̃ ρτ2B + τ3B .


(128)

Solving for η gives equation (25).

Derivation of equation (26). Differencing the Euler equations for periods 4 versus 5 for the Bonus
and Bonus Control groups gives

η x4B − x4BC + ζ sB4 − sBC = (1 − α)δ ρ (η − ζ ) x̃5B − x̃5BC + ζ s̃B5 − s̃BC




4 5 (129)

Habit stock evolution implies sB4 −sBC B BC B BC = ρ 2 τ B +ρτ B and s̃B − s̃BC = ρ sB − sBC + xB − xBC =



4 = ρ s3 − s3 + x3 − x3 2 3 5 4 4 4 4

5
ρ ρ 2 τ2B + ρτ3B + τ4B . Linear policy functions imply x̃5 = λ̃ s̃5 + µ̃5 , so x̃5B − x̃5BC = λ̃ s̃B5 − s̃BC


5 . Substituting
these equations gives

h 
i
ητ4B + ζ ρ 2 τ2B + ρτ3B = (1 − α)δ ρ (η − ζ )λ̃ + ζ ρ ρ 2 τ2B + ρτ3B + τ4B


(130)
h  
i
= (1 − α)δ ρ 2 ηλ + ζ 1 − λ̃ ρ 2 τ2B + ρτ3B + τ4B . (131)

Collecting ζ terms gives

h  
i
ζ ρτ3B + ρ 2 τ2B −(1−α)δ ρ 2 ζ 1 − λ̃ ρ 2 τ2B + ρτ3B + τ4B = −ητ4B +(1−α)δ ρ 2 η λ̃ ρ 2 τ2B + ρτ3B + τ4B .

(132)
Solving for ζ gives equation (26).

System of equations for (1 − α), η, and ζ .

First, we solve explicitly for (1 − α) before substituting it back in Equations (25) and (26) to solve for
η and ζ .

120
Online Appendix Allcott, Gentzkow, and Song

We define
−τ4B + (1 − α)δ ρ 2 λ ρ 2 τ2B + ρτ3B + τ4B
 
y := B  . (133)
ρτ3 + ρ 2 τ2B − (1 − α)δ ρ 2 (1 − λ ) ρ 2 τ2B + ρτ3B + τ4B
Observe that
ζ = η · y. (134)

We can use this observation to rearrange Equation (25):

pB − ζ ρτ2B + (1 − α)δ ρ 2 ζ (1 − λ ) ρτ2B + τ3B

η=
(135)
τ3B − (1 − α)δ ρ 2 λ ρτ2B + τ3B
η τ3B − (1 − α)δ ρ 2 λ ρτ2B + τ3B = pB − ζ ρτ2B − (1 − α)δ ρ 2 (1 − λ ) ρτ2B + τ3B




(136)
= pB − η · y ρτ2B − (1 − α)δ ρ 2 (1 − λ ) ρτ2B + τ3 B



(137)
pB = η τ3B − (1 − α)δ ρ 2 λ ρτ2B + τ3B + y ρτ2B − (1 − α)δ ρ 2 (1 − λ ) ρτ2B + τ3B .





(138)

Then, define

x := τ3B − (1 − α)δ ρ 2 λ ρτ2B + τ3B + y ρτ2B − (1 − α)δ ρ 2 (1 − λ ) ρτ2B + τ3B

(139)

where we observe that

pB
η= , (140)
x
and
pB y
ζ= . (141)
x
Finally, we get that

ητ2B
(1 − α) =   (142)
δ ρ −pB + (η − ζ )τ̃3B + ζ ρτ2B
pB B
x τ2
= h B B B
i. (143)
δ ρ −pB + ( px − px y )τ̃3B + px y ρτ2B

Since all scalars are known in the last equation, we can now solve for α. Then, we can estimate η and
ζ by substituting α in Equations (25) and (26) respectively.

121
Online Appendix Allcott, Gentzkow, and Song

G.3 Naivete about Temptation

Derivation of equation (27). The Euler equation predicted for period t on the survey at the beginning of
period t is
h i
ηxt∗ (st , γ̃, pt ) + ζ st + ξt − pt + γ̃ = (1 − α)δ ρ η x̃t+1 + ζ st+1 + ξt+1 − pt+1 + γ̃ + γ̃ λ̃ − (ζ x̃t+1 + φ ) .
(144)
This equation uses the assumption that consumers are aware of period t projection bias when predicting
period t consumption on survey t, so the only reason why the period t survey-taker mispredicts the period t
objective function is naivete about period t temptation.
Habit stock evolution implies s̃t+1 = ρ(st + x̃t ). Linear policy functions imply x̃t+1 = λ̃ s̃t+1 + µ̃t+1 .
Substituting these equations into the predicted Euler equation gives
h   i
ηxt∗ (st , γ̃, pt ) + ζ st + ξt − pt + γ̃ = (1 − α)δ ρ (η − ζ ) λ̃ s̃t+1 + µ̃t+1 + ζ s̃t+1 + ξt+1 − pt+1 + γ̃ + γ̃ λ̃ − φ .
(145)
h  i
= (1 − α)δ ρ (η − ζ )λ̃ + ζ s̃t+1 + (η − ζ )µ̃t+1 + ξt+1 − pt+1 + γ̃ + γ̃ λ̃ − φ
(146)
h  i
= (1 − α)δ ρ (η − ζ )λ̃ + ζ ρ(st + x̃t ) + (η − ζ )µ̃t+1 + ξt+1 − pt+1 + γ̃ + γ̃ λ̃ − φ .
(147)

Analogously, the actual Euler equation for period t can be written as

h  i
ηxt∗ (st , γ, pt )+ζ st +ξt − pt +γ = (1−α)δ ρ (η − ζ )λ̃ + ζ ρ(st + xt∗ ) + (η − ζ )µ̃t+1 + ξt+1 − pt+1 + γ̃ + γ̃ λ̃ − φ .
(148)
Differencing the actual and predicted Euler equations for period t versus period t + 1 for the Control
group gives
h 
i
η xCt − x̃Ct + γ − γ̃ = (1 − α)δ ρ (η − ζ )λ̃ + ζ ρ xCt − x̃Ct


(149)

Solving for γ − γ̃ and substituting mC = xCt − x̃Ct gives equation (27).

G.4 Temptation
Limit effect: derivation of equation (28). Consider a “zero temptation” intervention that fully eliminates
both perceived and actual temptation starting in period 2, generating treatment effects τt0 . Differencing the
average Euler equations for periods 2 versus 3 for the zero temptation group versus its control group gives

122
Online Appendix Allcott, Gentzkow, and Song

h i
η x20 − x20C − γ = (1 − α)δ ρ (η − ζ ) x̃30 − x̃30C + ζ s̃03 − s̃0C




3 − γ̃ − γ̃ λ̃ (150)
h i
ητ20 − γ = (1 − α)δ ρ (η − ζ )τ̃30 + ζ ρτ20 − γ̃ − γ̃ λ̃ (151)

Solving for γ and substituting τ 0 = τ L /ω gives equation (28).

To solve for γ as a function of data and known parameters, we solve equation (27) for γ̃, substitute into
equation (28), and rearrange, giving

 h  i
ητ2L /ω − (1 − α)δ ρ (η − ζ )τ̃3L /ω + ζ ρτ2L /ω + (1 + λ̃ )mC2 · −η + (1 − α)δ ρ 2 (η − ζ )λ̃ + ζ

γ= .
1 − (1 − α)δ ρ(1 + λ̃ )
(152)

Bonus valuation: derivation of equation (29). When we elicited the bonus valuation on survey
2, we had not yet told participants whether the bonus would be in effect for period 2 or 3. The theoretical
valuations for a period 2 vs. period 3 bonus are identical if we assume that consumers predict no anticipatory
effect of the period 3 bonus. Otherwise, this derivation would need to account for the period 2 survey taker’s
valuation of the perceived internality reduction from the anticipatory effect. Since the actual bonus was for
period 3, we focus the derivation on that case and maintain the assumption of zero predicted anticipatory
effect.
From the perspective of the period 2 survey taker, the predicted period 3 continuation value (given
naivete about future projection bias) as a function of predicted habit stock and period 3 price is

V3 (s̃3 , p3 ) =u3 (x̃3∗ (s̃3 , γ̃, p 3 ) ; s̃3 , p3 ) + δV4 (s̃4 , ·) . (153)

The change in that predicted continuation value from a marginal change in period 3 price is

 
dV3 (s̃3 , p3 ) ∂ ũ3 ∂ x̃3 ∂ ũ3 dV4 (s̃4 , ·) ∂ s̃4
= + +δ . (154)
d p3 ∂ p3 ∂ p3 ∂ x̃3 d s̃4 ∂ x̃3

People taking survey 2 predict that their period 3 selves will set x3 to maximize that same function with
an additional γ̃x3 in period 3 flow utility:

x̃3∗ (s̃3 , γ̃, p 3 ) = arg maxu3 (x3 ; s̃3 , p 3 ) + γ̃x3 + δV4 (s̃4 , ·) . (155)
x3

Thus, people taking survey 2 predict that they will set x3 such that

123
Online Appendix Allcott, Gentzkow, and Song

∂ ũ3 dV4 (s̃4 , ·) ∂ s̃4

+ γ̃ + δ = 0. (156)
∂ x3 d s̃4 ∂ x̃3

Substituting equation (156) into equation (154) gives

dV3 (s̃3 , p3 ) ∂ ũ3 ∂ x̃3

= − γ̃ (157)
d p3 ∂ p3 ∂ p3
∂ x̃3
= −x̃3 (p3 ) − γ̃ . (158)
∂ p3

This illustrates a temptation-adjusted envelope theorem: the effect of a marginal price change on the
long-run self’s utility (given perceived misoptimization from the long-run self’s perspective) equals the
mechanical effect x̃3 (p3 ) adjusted by the magnitude of the perceived misoptimization γ̃ ∂∂ x̃p33 . With zero
perceived temptation (γ̃ = 0), this reduces to the standard envelope theorem. The derivation for a period
2 bonus would be analogous, except with (1 − α) multiplying the predicted period 3 continuation value in
both the survey taker’s objective function and the predicted period 2 objective function.
We integrate over equation (158) to determine the effect of a non-marginal price increase from 0 to pB :

Z p3 =pB3 ∂ x̃3
V3 s̃3 , p3 = pB3 −V3 (s̃3 , p3 = 0) =


−x̃3 (p3 ) − γ̃ d p3 (159)
p3 =0 ∂ p3
= −pB3 · x̃3 (pB3 ) + x̃3 (0) /2 − γ̃ · x̃3 (pB3 ) − x̃3 (0) ,



(160)

where the second line follows from the fact that demand is linear in price, which was shown in Proposition
2.

Limit valuation: derivation of equation (31). The period 3 survey-taker’s objective function is

α ∑Tr=4 δ r−3 ur (x̃r∗ (s3 , γ̃, p r ) ; s3 , pr )

V3 (s3 , γ̃3 ) = u3 (x3∗ (s3 , γ̃3 , p 3 ) ; s3 , p3 ) + . (161)
+(1 − α)δV4 (s̃4 , ·)
This equation uses the assumption that the survey taker is projection biased.
The change in that objective function from a marginal change in perceived period 3 temptation is

dV3 (s3 , γ̃3 ) ∂ x3∗ (s3 , γ̃3 , p 3 ) ∂ u3

 
∂V4 (s̃4 , ·) ∂ s̃4
= + (1 − α)δ . (162)
d γ̃3 ∂ γ̃3 ∂ x3 ∂ s̃4 ∂ x̃3
People taking survey 3 predict that they will set x3∗ such that

124
Online Appendix Allcott, Gentzkow, and Song

∂ u3 dV4 (s̃4 , ·) ∂ s̃4

+ γ̃3 + (1 − α)δ = 0. (163)
∂ x3 d s̃4 ∂ x̃3

Substituting the period 3 first-order condition from equation (163) into equation (162) gives

dV3 (s3 , γ̃3 ) ∂ x∗ (s3 , γ̃3 , p 3 )

= −γ̃3 3 . (164)
d γ̃3 ∂ γ̃3
We integrate over equation (164) to determine the effect of the non-marginal temptation reduction from
γ̃ to (1 − ω)γ̃:

∂ x3∗ (γ̃3 )
Z γ̃3 =(1−ω)γ̃
L
v = V3 (s3 , γ̃3 = (1 − ω)γ̃) −V3 (s3 , γ̃3 = γ̃) = −γ̃3 d γ̃3 . (165)
γ̃3 =γ̃ ∂ γ̃3
1 1
= (x3∗ (γ̃) − x3∗ (0)) · γ̃ · − (1 − ω)2 · (x̃3 (γ̃) − x̃3 (0)) · γ̃ ·
2 2
(166)

1
= (x3∗ (γ̃) − x3∗ (0)) · γ̃ · 1 − (1 − ω)2 · (167)
2
ω(2 − ω)
= (x3∗ (γ̃) − x3∗ (0)) · γ̃ · (168)
2
∗ ∗
(x (γ̃) − x3 ((1 − ω)γ̃)) ω(2 − ω)
= 3 · γ̃ · (169)
ω 2
(2 − ω)
= (x3∗ (γ̃) − x3∗ ((1 − ω)γ̃)) · γ̃ · , (170)
2

where the second line is the area of the long-run self’s perceived deadweight loss reduction trapezoid (fol-
lowing from linear demand) and the fifth line follows from the assumption that τ̃ L /ω = τ̃ 0 .

G.5 Temptation with Multiple Goods

We now extend our model to include a second temptation good y, which in our experiment is FITSBY use
on other devices. Habit stock now evolves according to st+1 = ρ (st + xt + yt ). Before period t, consumers
now consider flow utility to be ut (xt , yt ; st , pt ). In period t, consumers choose as if period t flow utility is
ut (xt , yt ; st , pt ) + γx xt + γy yt . Before period t, consumers predict that they will choose as if period t flow
utility is ut (xt , yt ; st , pt ) + γ̃x xt + γ̃y yt . x is still sold at price pt , while yt has zero price. The limit treatment
fully eliminates perceived and actual temptation on x.
We derive new equations for γ or γ̃ for the limit effect, bonus valuation, and limit valuation strategies.
With all three strategies, if y is not a temptation good (γ̃y = γy = 0) or if y is neither a substitute nor a
complement for x, then our original estimating equations are unaffected.
Limit effect. To derive γ using the limit effect strategy, we assume full projection bias (α = 1). We

125
Online Appendix Allcott, Gentzkow, and Song

assume that the static quadratic flow utility function is now

ηx 2 ηy
u(x, y; p) = x + ξx x − px + σ xy + y2 + ξy y. (171)
2 2
Without the limit, consumers maximize u(x, y; p) + γx x + γy y, giving

σ x + ξy + γy
y∗ (x) = (172)
−ηy
ξy +γy
ξx − p + σ −ηy + γx
∗
x = 2 (173)
−ηx + σηy

Taking the expectation over individuals, the bonus effect on x∗ is

pB
τxB = 2 (174)
−ηx + σηy

The limit allows consumers to set xL before period t. When setting the limit, consumers predict that in
period t they will set y conditional on xL to maximize u (xL , y; p) + γ̃x xL + γ̃y y, giving

σ xL + ξy + γ̃y
y∗ (xL ) = . (175)
−ηy
Consumers thus set xL to maximize u (xL , y∗ (xL ) ; p), giving
σ
ξx − p + ξy −η y
xL = 2 . (176)
−ηx + σηy
σ xL +ξy +γ̃y σ x∗ +ξy +γy
The effect of the limit on y is y∗ (xL ) − y∗ (x∗ ) = −ηy − −ηy . Taking the expectation over
individuals, the limit effect on y is

σ L
τyL = τ (177)
−ηy x

The effect of the limit on x is

126
Online Appendix Allcott, Gentzkow, and Song

σ ξy +γy
ξx − p + ξy −η ξx − p + σ −ηy + γx
∗ y
xL − x = 2 − 2 (178)
−ηx + σηy −ηx + σηy
σ
−ηy γy + γx
= 2 (179)
−ηx + σηy
 
σ
−γ 1 + −η y
= 2 (180)
−ηx + σηy

where the third line assumes γx = γy = γ.

Taking the expectation over individuals and substituting equations (174) and (177) gives

τL
 
−γ 1 + τyL
τxL = x
. (181)
pB /τxB
Rearranging gives

τxL · pB /τxB


γ= τL
. (182)
1 + τyL
x

This exactly parallels equation (28) for the α = 1 case, except adjusting the denominator for substitution. If
x and y are substitutes, then the estimated γ increases: more temptation is required to explain a given limit
when the consumer knows that she can evade the limit through substitution to another temptation good. If
x and y are complements, then the estimated γ decreases: less temptation is needed to explain a given limit
when the consumer knows that the limit will also cause reductions in another temptation good.
Bonus valuation. The derivation for the bonus valuation with substitute goods is very similar to the
one-good case. The change in the period 3 continuation value function from a marginal change in p3 is

   
dV3 (s̃3 , p3 ) ∂ ũ3 ∂ x̃3 ∂ ũ3 dV4 (s̃4 , ·) ∂ s̃4 ∂ ỹ3 ∂ ũ3 dV4 (s̃4 , ·) ∂ s̃4
= + +δ + +δ . (183)
d p3 ∂ p3 ∂ p3 ∂ x̃3 d s̃4 ∂ x̃3 ∂ p3 ∂ ỹ3 d s̃4 ∂ ỹ3

People taking survey 2 predict that they will set x3 and y3 according to

∂ ũ3 dV4 (s̃4 , ·) ∂ s̃4

+ γ̃x + δ =0 (184)
∂ x3 d s̃4 ∂ x̃3
∂ ũ3 dV4 (s̃4 , ·) ∂ s̃4
+ γ̃y + δ = 0. (185)
∂ y3 d s̃4 ∂ ỹ3

∂ ũ3
Substituting equations (184) and (185) as well as ∂ p3 = −x̃3 (p3 ) into equation (183) gives

127
Online Appendix Allcott, Gentzkow, and Song

dV3 (s̃3 , p3 ) ∂ x̃3 ∂ ỹ3

= −x̃3 (p3 ) − γ̃x − γ̃y . (186)
d p3 ∂ p3 ∂ p3

Integrating over a non-marginal price increase from 0 to pB assuming linear demand, also assuming
γ̃x = γ̃y = γ̃, taking the expectation over participants, and rearranging gives

v̄B − F̄ B + pB3 x̃¯3B+BC

γ̃ =   (187)
− τ̃x3B + τ̃ B
y3

This exactly parallels equation (30), except

 adjusting the denominator for substitution. The survey taker
B B
values the total temptation reduction − τ̃x3 + τ̃y3 induced by the bonus. If x and y are substitutes, the
total temptation reduction is lower, and more temptation is needed to justify a given valuation. If x and y
are complements, the total temptation reduction is higher, and less temptation is needed to justify a given
valuation.
Limit valuation. The derivation for the limit valuation with substitute goods is also similar to the one-
good case. The change in the period 3 survey-taker’s objective function from a marginal change in perceived
period 3 temptation for good x only is

dV3 (s3 , γ̃x3 ) ∂ x3∗ ∂ u3 ∂ y∗ ∂ u3

   
∂V4 (s̃4 , ·) ∂ s̃4 ∂V4 (s̃4 , ·) ∂ s̃4
= + (1 − α)δ + 3 + (1 − α)δ . (188)
d γ̃x3 ∂ γ̃x3 ∂ x3 ∂ s̃4 ∂ x̃3 ∂ γ̃x3 ∂ y3 ∂ s̃4 ∂ ỹ3

Substituting the predicted period 3 first-order conditions for x and y gives

dV3 (s3 , γ̃x3 ) ∂ x∗ ∂ y∗

= −γ̃x3 3 − γ̃y 3 . (189)
d γ̃x3 ∂ γ̃x3 ∂ γ̃x3
Integrating over this from γ̃x to (1 − ω)γ̃x assuming linear demand, also assuming γ̃x = γ̃y = γ̃, taking
the expectation over participants, and rearranging gives

v̄L
γ̃ =  . (190)
− τ̃3L (2 − ω)/2 + τ̃y3
L

As with the bonus valuation, the survey taker values the total temptation deadweight loss reduction induced
by the limit. If x and y are substitutes, the total temptation reduction is lower, and more temptation is needed
to justify a given valuation. If x and y are complements, the total temptation reduction is higher, and less
temptation is needed to justify a given valuation.

G.6 Intercept
Derivation of equation (33).

128
Online Appendix Allcott, Gentzkow, and Song

Re-arranging steady state consumption from equation (21) gives

h   i
(1 − α)δ ρ(φ − ξ ) + ξ − (1 − (1 − α)δ ρ) p + (1 − α)δ ρ (ζ − η)mss − 1 + λ̃ γ̃ + γ =

ρ − (1 − α)δ ρ 2
 (191)
xss −η − (1 − α)δ ρ(ζ − η) − ζ .
1−ρ

Solving for the intercept and substituting xi1 = xss gives equation (33).

H Counterfactual Simulations Appendix

Table A14: Effects of Temptation and Habit Formation on FITSBY Use

(1) (2)
Restricted Unrestricted
model model
FITSBY use (minutes/day) B
(τ2 = 0, α = 1) (α = α̂)
Baseline 153 153
[149, 157] [149, 157]
No naivete 153 151
[149, 157] [140, 156]
No temptation 105 103
[76.9, 120] [67.0, 119]
No habit formation 78.1 73.3
[50.2, 102] [43.1, 99.4]
No temptation or habit formation 53.8 49.0
[25.6, 77.9] [17.2, 75.7]
Notes: This table presents point estimates and bootstrapped 95 percent confidence intervals for predicted steady-state
FITSBY use with different parameter assumptions, using equation (15). The numbers are as plotted in Figure 10.

129
Online Appendix Allcott, Gentzkow, and Song

Table A15: Effects of Temptation on FITSBY Use Under Alternative Assumptions

(1) (2)
Restricted Unrestricted
model model
Effect of temptation on FITSBY use (minutes/day) B
(τ2 = 0, α = 1) (α = α̂)
Limit effect 47.5 49.5
[34.3, 75.0] [34.9, 86.4]
Bonus valuation 70.5 71.2
[49.2, 116] [49.5, 118]
Limit valuation 61.5 62.3
[42.8, 103] [43.3, 106]
Limit effect, multiple-good model 57.4
[40.0, 97.0]
Bonus valuation, multiple-good model 63.2 63.9
[44.5, 103] [44.7, 107]
Limit valuation, multiple-good model 91.3 91.6
[50.1, 155] [51.0, 155]
Limit effect, ω = ω̂ 123 127
[85.2, 155] [87.3, 156]
Limit valuation, ω = ω̂ 42.7 43.8
[29.7, 71.1] [30.2, 76.6]
Heterogeneous limit effect 47.1 48.6
[34.2, 71.9] [34.6, 76.5]
Limit effect, weighted sample 52.2 57.8
[32.3, 112] [33.9, 144]
Notes: This table presents point estimates and bootstrapped 95 percent confidence intervals for the effects of temp-
tation on average steady-state FITSBY use, using equation (15). The first nine estimates are for the nine temptation
estimation strategies presented in Table A9. The tenth estimate is for the limit effect strategy after reweighting the
sample to be more representative of U.S. adults. Appendix Tables A11–A13 present the demographics, moments, and
parameter estimates in the weighted sample. Average baseline FITSBY use is 153 and 156 minutes per day for the un-
weighted and weighted samples, respectively. We do not have a limit effect estimate for the unrestricted multiple-good
model.

130
Online Appendix Allcott, Gentzkow, and Song

Figure A35: Distribution of Effects of Temptation on FITSBY Use

0.09
Fraction of sample

0.06

0.03

0.00

0 100 200 300

Effect of temptation on FITSBY use (minutes/day)
Notes: Using the heterogeneous limit effect strategy, we estimate temptation γ̂i for each Limit group participant, which
we then insert into equation (15) to predict the individual-specific effect of temptation on steady-state FITSBY use.
This figure presents the distribution of effects across participants, winsorized at 300 minutes per day.

131